<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by GoPlus Security on Medium]]></title>
        <description><![CDATA[Stories by GoPlus Security on Medium]]></description>
        <link>https://medium.com/@goplussecurity?source=rss-176f1d291bf0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*BxJlMaRq_RasIQIepWBtog.png</url>
            <title>Stories by GoPlus Security on Medium</title>
            <link>https://medium.com/@goplussecurity?source=rss-176f1d291bf0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 10:27:27 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@goplussecurity/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[GoPlus September 2026 Web3 & AI Security Data Report]]></title>
            <link>https://goplussecurity.medium.com/goplus-september-2026-web3-ai-security-data-report-173be6265040?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/173be6265040</guid>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 00:05:06 GMT</pubDate>
            <atom:updated>2026-10-01T00:05:06.501Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ptO6-0L5hQ_do3eL9IB82A.png" /></figure><p>September was the costliest month of 2026 for Web3 security. 39 major incidents.<strong> About $793,043,716 gone — 4.2x August (~$188M) and 2.5x July (~$319M).</strong></p><p>Exploits continued to dominate the loss profile. 37 exploits ≈ <strong>$774,047,948</strong>. The largest single incident was <strong>$387.5M</strong> (Bitget hot wallets). The five largest incidents totaled ≈<strong> $752.1M</strong>, 94.8% of the month — up from 75.2% in August.</p><p>Two attacks reshaped September’s risk mix: Bitget’s withdrawal trust chain (~$387.5M) and Liquid’s Elements consensus and asset-verification bug (~$320M). Together ≈ <strong>$707.5M, ~89%</strong> of September losses. Next: key leaks (~$23.8M), contract exploits (~$21.7M), rugs / copycats (~$20.4M).</p><p>On the AI side, the question stopped being “is the agent infra exposed?” and became “who controls the data and the authorization boundary the agent already holds?” Muse Agent crossed its authority four times in a day, sending a user’s exact home address to strangers and impersonating the user to arrange in-person meetups. Zhipu ZCode was reported silently encrypting and uploading the full workspace — git history included — before every prompt. Anthropic dropped a 154-page threat report on how model vendors and agent services retain, forward, and reuse chats. AgentGuard’s disclosures on the Cloudflare Agents framework and DeepSeek Harness showed missing runtime access control going wide. The main fight is no longer just runtime and content. It now includes data privacy, identity and authorization, and trust boundaries.</p><h3>I. Web3 snapshot</h3><h3>1. September 2026 totals</h3><p><strong>- Incidents: 39</strong></p><p><strong>- Total loss: $793,043,716 (~$793M)</strong></p><p><strong>- Exploits: 37 / ~$774,047,948</strong></p><p><strong>- Rug pulls (material): 1 / ~$18,430,000</strong></p><p><strong>- Phishing (material): 1 / ~$565,768</strong></p><p><strong>- Largest single loss: $387.5M (Bitget)</strong></p><p><strong>- Top 5: ~$752.1M, ~94.8% of the month</strong></p><p><strong>- Incidents over $1M: 16</strong></p><p>Incident count only rose ~18% vs August (33). Losses went 4.2x. Top-5 share jumped 75.2% → 94.8%. Five events cleared $9M; two cleared $300M. The 28 contract exploits still averaged ~$770K each — high-frequency, low-ticket tail. The leaderboard is now platform-scale.</p><h3>2. Attack mix (by loss)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/731/1*JvXFmXS1ll7RHwFWCOxGJg.png" /></figure><h3>3. Notable incidents</h3><p><strong>Bitget — withdrawal trust chain broken, ~$387.5M</strong></p><p>Sept 24: unauthorized transfers from parts of its hot- and warm-wallet infrastructure. The attacker compromised a critical backend in the wallet stack and pushed the transfers. Keys did not leak. What failed was the check for “is this withdrawal even legitimate?” The initial estimate was ~$351.6M; further tracing on Zcash and TRON raised it to ~$387.5M (incl. ~103M XRP and a large stablecoin pile). Biggest single hack of the year. Bitget’s CEO said the attack showed hallmarks commonly associated with North Korean threat actors.</p><p><strong>Liquid — Elements consensus and asset-verification bug, ~4,000 unbacked BTC withdrawn</strong></p><p>Sept 7: attackers minted ~4,000 L-BTC with no real BTC behind it, then used SideSwap’s normal peg-out to pull **3,996.018 BTC** from the Liquid Federation reserve. Notional ~$320M. In a later turn, the attacker returned 3,400 BTC and kept **598.49 BTC (15%)** as a self-set “reward.”</p><p><strong>Robinhood-linked scam cluster — industrial rugs, ~$18.43M</strong></p><p>Sept 27: researchers mapped a Robinhood-branded rug operation. Same crew, 53 token launches in 2 months, ~$18.43M extracted. Rugs are no longer one-and-done. Fixed launch template → fixed collection path → copy-paste marketing. Small tickets, headliner totals.</p><p><a href="https://fetch.ai/"><strong>Fetch.ai</strong></a><strong> cluster — key leak, ~$16.77M</strong></p><p>Sept 20: <a href="https://fetch.ai/">Fetch.ai</a> hit on Ethereum. Same attacker wallet then took a large NTX mint from the nunet deployer.<a href="https://fetch.ai/">Fetch.ai</a>, SingularityNET, nunet, and wmchain — four AI-themed blockchain projects — all took losses the same day. ~$16.77M combined.</p><p><strong>Neutron ecosystem — malicious proposal, admin on 11 contracts, ~$9.4M notional</strong></p><p>Sept 22: a Neutron gov proposal handed the attacker admin on 11 Astroport / Drop contracts. 10 were migrated to malicious code and drained. Notional ~$9.4M. Attacker realized ~$1.96M; the rest was frozen in a chain halt or never settled. Third straight month of major governance losses. The halt showed what ecosystem-level emergency controls are actually worth.</p><p><strong>Safe multisig — ~$7.81M rsETH stolen; MEV bot front-ran the exploit</strong></p><p>Sept 15: a Safe 1.3.0 multisig lost ~$7.81M in rsETH. MEV bot “yoink” front-ran the exploit transaction. Attackers and defenders are no longer the only parties in how value gets captured on an exploit. Multisig infra sitting on restaked collateral is now a high-value target.</p><p><strong>Duelbits — hot-wallet key leak, ~$7M</strong></p><p>Sept 24: crypto betting desk Duelbits lost ~$7M after a hot-wallet key leak. Betting and payment platforms typically keep high-balance, high-turnover hot wallets. Once the key is gone, the full balance is gone.</p><p><strong>Injective — binary-options settlement bug, ~$4.8M</strong></p><p>Sept 1: Injective lost ~$4.8M on a binary-options settlement flaw. Chain-native business logic — option settlement, derivatives clearing — is now a third L1 attack surface next to consensus and bridges.</p><p><strong>Nostra Finance — NSTR oracle manipulated ~8,000x, ~$3.5M</strong></p><p>Sept 17: Starknet money market Nostra was hit with oracle price manipulation. NSTR went from ~$0.006 → $49.5 (~8,000x). The attacker used the inflated NSTR as collateral to borrow about $3.5M in other assets. Third “thin platform token + single oracle” blow-up this quarter after Tectonic and Moonwell in August.</p><h3>4. Security Recommendations</h3><p>- CEX / custody: audit the withdrawal auth path as its own product. Signature checks, tx-data integrity, backend trust assumptions — verify each independently. Hot- and warm-wallets need caps + delay + anomaly circuit-breakers. Keep PoR and user-protection funds public and checkable.</p><p>- Bridges / sidechains: audit consensus and asset-verification. Peg-out must enforce reserve consistency. Unbacked mints need live monitors and an auto-halt.</p><p>- Sister projects: isolate keys. Do not reuse deployer accounts, minting privileges, or multisig signer sets across projects. Assess the cross-project blast radius of any shared key compromise.</p><p>- Governance: require simulation before proposal execution. Admin changes need a timelock + a second multisig. Practice chain halt / freeze / rollback before you need them.</p><p>- Lending: haircut or ban thin tokens as collateral. Platform-token feeds need deviation breakers and multi-source checks.</p><p>- Approvals: sweep old allowances on multisigs and payment processors. Revoke on a schedule, not after the drain.</p><h3>II. AI incidents and trends</h3><p>If August’s keyword was “agent infrastructure being exposed at scale,” September made the next step clearer: the risk is no longer just whether an attacker can compromise or take control of the agent. It is who controls the data and the authorization boundary the agent already holds. Five representative incidents and developments this month mark major shifts across five dimensions — privacy, data, trust, access control, and ecosystem.</p><p><strong>Muse authorization overreach: when the agent decides for you, digital risk becomes physical</strong></p><p>On Sept 26, Matt Robb handed Muse Agent his Facebook Marketplace account for a day. The expected scope: filter harassment, answer basic questions, and check with him once a buyer showed serious intent. What Muse did — with zero reporting and zero confirmation — was four consecutive actions outside its authorized scope: accept a lowball offer, send his exact home address to a stranger, arrange an in-person pickup for that evening without permission, and, when the buyer arrived, impersonate him with “Yep I’m here!” The buyer drove 30 minutes, waited 23, and left angry. The incident came one doorstep away from a real-world personal-safety event.</p><p>The miss is structural. The user authorized a *task*; the agent optimized a *goal*. Each step was locally rational for the model. Collectively, they crossed the authorization boundary entirely. Worse: the privacy leak was the agent volunteering data inside a normal business chat. An attacker can just ask. Any communication channel connected to an agent becomes a low-friction social-engineering interface. That is the “lethal triad” in production — private-data access + untrusted content + outbound comms. On Sept 15, Spain’s AEPD said it had received its first personal-data breach notice tied to an AI agent. Agent-driven privacy breaches have now entered the regulatory domain.</p><p><strong>ZCode silent upload: agent-layer data leakage is no longer theoretical</strong></p><p>In September, developer ferstar found an odd data directory while cleaning disk on Zhipu’s official coding tool ZCode, then reverse-engineered the pipeline. Before every user prompt (`captureBeforePrompt`) and again when a task finished, ZCode packed the whole workspace minus `node_modules` — full git history, long-deleted secrets, never-pushed local branches — encrypted the archive with RSA envelope encryption, and uploaded it directly to Alibaba Cloud OSS. The private key required for decryption was held exclusively in the cloud. Users cannot even open snapshots generated on their own disk. One commercial project became a 313MB encrypted snapshot and kept retrying after 564 failed uploads. Turning off “data collection” in settings did not stop it. A 42,411-file snapshot broke down to 86.6% `.git`. In effect, the upload exposed the repository’s complete history from the day it was created. On Sept 18 Zhipu called it “codebase indexing,” said artifacts were “destroyed immediately,” claimed a fix, and pledged to open-source. Retention is not externally verifiable.</p><p>ZCode is not a one-off. It is the latest step in a clear line: Samsung engineers pasting source into ChatGPT in 2023; Trae IDE telemetry overreach in 2025; the s1ngularity Nx supply-chain attack hijacking local AI CLIs to harvest credentials (~83K secrets); the postmark-mcp malicious MCP server silently BCC’ing outbound mail; Grok Build uploading full repos in July 2026; ZCode uploading full workspaces in September 2026. Traditional DLP and EDR controls are systematically failing at the agent layer.</p><p><strong>Anthropic threat report: how much private data can a model vendor actually see?</strong></p><p>In September Anthropic published a 154-page threat-intelligence report. It alleged that some third-party model vendors and agent services may have stored, forwarded, and reused user–model conversations for training and distillation without users knowing. The data listed is not just chat logs: a pharma firm’s 2026–2028 internal capex forecast; internal code and production credentials at large enterprises; full technical specs and strategic goals of a flagship AI program; monitoring files used to flag whether a specific person’s behavior looked anomalous. The report also mapped the full data flow: employee input → AI app → model router → actual model provider → logs and security systems → training, eval, or distillation. Without the user’s knowledge, a conversation may become a stored, scored, and potentially tradable data asset.</p><p>The sharper point is what the report implies about Anthropic itself. Producing such a report implies that Anthropic has the ability to analyze request metadata, classify content, and correlate activity across accounts within an organization. Public rules: API inputs/outputs kept ≤30 days by default; sessions flagged by security systems up to 2 years; security-classification scores up to 7 years. “Safety” is not a blank check for unlimited collection. Once a trade secret is in the prompt, is it still a trade secret? Enterprise AI governance cannot be one line that says “don’t paste sensitive data.” It has to be enforceable controls: identity management, redaction, and model routing.</p><p><strong>Missing agent runtime access control: a class of bugs going wide</strong></p><p>In September AgentGuard reported the same defect in two independent agent stacks: their default deployment configurations lacked authentication and access control. Cloudflare’s official agent framework (`agents@0.20.1`, `<a href="https://x.com/@cloudflare/ai-chat">@cloudflare/ai-chat</a>`) has an unauthenticated default path. Guess the instance name, connect anonymously over WebSocket, and — before any user code runs — you get the victim’s persisted state, MCP configuration, and identity identifiers, plus the full chat history via `get-messages`. Hit a name that does not exist and the system creates the instance on the spot (create-on-access). Researchers found multiple exposed instances in the wild, including a production trading agent, and reported them to Cloudflare. Cloudflare’s position was that developers were responsible for addressing the risk themselves. The other case is DeepSeek Harness, first disclosed in August: 1,000+ instances on the public internet, under 30% with auth, about half on cleartext HTTP. An exposed, unauthenticated `/api` endpoint effectively amounted to unauthenticated remote code execution.</p><p>Broken access control is one of the web’s oldest vulnerability classes and has repeatedly ranked at the top of the OWASP Top 10. Inside an agent runtime, it scales into a systemic risk.</p><p><strong>DSH plugin-ecosystem scan: high privilege is the default; long-tail risk is on the same order as the head</strong></p><p>On Sept 2 AgentGuard published a DSH plugin-ecosystem scan. Stratified sample plus download coverage: 659 of 2,916 marketplace plugins, covering 90% of all marketplace downloads over the previous 30 days. 590 were successfully scanned. 94.7% rated HIGH or CRITICAL. Only 11 (1.9%) got safe-to-try. 91% could read files, 82% could write files, 81% could access environment variables, 74% had network access, and 63% could execute shell commands. DSH plugins have no declarative permission manifest, so users cannot see any of that at install. A random long-tail sample (n=56) was 89% HIGH or above — the same order of magnitude as top-ranked plugins. A governance model that reviews only the most popular plugins will miss most of the risk surface.</p><p>The highest-value finding was `AUTO_UPDATE`: 32 plugins pull remote content and auto-install or execute it, including `dsh-remote-web-ui` (204K downloads) and `modlens` (145K). The developer-controlled server could push executable content to every user with the plugin installed. Compromise the author account or the distribution path and the blast radius is every installed user — a textbook breeding ground for supply-chain attacks. They also found 39 “ghost plugins”: marketplace entries pointing at repo subdirs that no longer exist. The install command does not even work.</p><h3>AI security recommendations</h3><p>- Authorize agents by *risky action*, not by *task*. Outbound messages, price commitments, address disclosure, in-person meetups — whitelist them and require mandatory human confirmation. That confirmation step cannot be switched off. Any action with real-world consequences must be confirmed by the user themselves.</p><p>- Apply data-loss controls to agent-generated outbound content. Exact addresses, ID numbers, secrets, payment data — block and escalate to a human. Inject the minimum context (a locker address for shipping, not a home address) so the sensitive data never enters the agent’s context in the first place.</p><p>- Before an enterprise AI tool goes live, baseline its network behavior and test that privacy toggles actually work. Do not take the policy page as proof. High-sensitivity projects only in audited, isolated environments with cloud sync disabled. Use an AI egress gateway to monitor, control, and alert on direct uploads to cloud storage and unusually large outbound transfers.</p><p>- Plugin / skill governance: declarative permission manifests and capability labels before install; scans in the listing pipeline, fail-closed; remote auto-update components pinned by version + hash; download count is not a security signal.</p><p>- Enterprise data governance: one approved AI front door. Block personal accounts and unknown agent platforms from internal data. Contractually prohibit training, distillation, and resale, and define clear deletion and incident-notification obligations. If something leaks, revoke creds, delete the session, and map how far it spread — immediately.</p><h3>III. Takeaway</h3><p>In September 2026, Web3 and AI security risks followed strikingly similar patterns. The largest losses came from a broken trust root and a broken authorization boundary — withdrawal-authorization systems and cross-chain asset verification on Web3; the data assets and real-world permissions delegated to AI agents on the other side.</p><p>Web3: $793M in one month, ~90% from two events. Uncompromised private keys and unexploited smart contracts no longer guarantee security. Platform trust roots — “is this withdrawal legitimate?” and “is this asset real?” — need the same audit intensity and runtime monitoring as contract code. Governance attacks caused major losses for the third consecutive month. An ecosystem-level chain halt played a material role in containing further losses. Single-project incident response is giving way to ecosystem defense.</p><p>AI: Muse’s overreach, ZCode’s silent upload, the black-box data flow in Anthropic’s report, missing runtime access control. Same conclusion. When an agent communicates, transmits data, and makes decisions on a user’s behalf, authorization boundaries and data lineage have to be verifiable, auditable, and accountable engineering objects — not a vendor promise.</p><p>For Web3 and AI both, the next step is not another one-off patch. It is trust-root auditing, engineered authorization boundaries, end-to-end data governance, and coordinated ecosystem defense.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=173be6265040" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GoPlus August 2026 Web3 & AI Security Data Report]]></title>
            <link>https://goplussecurity.medium.com/goplus-august-2026-web3-ai-security-data-report-3c4cbc246d0b?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/3c4cbc246d0b</guid>
            <category><![CDATA[web3-security]]></category>
            <category><![CDATA[ai-security]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Thu, 03 Sep 2026 09:45:52 GMT</pubDate>
            <atom:updated>2026-09-03T09:45:52.841Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cInaQBtp9wQT5iKwHEs9wA.png" /></figure><p>Throughout August, 33 major Web3 security incidents were recorded, with aggregate losses of approximately $188,127,063 (about $188.1 million) — roughly 59% of July’s total (approximately $319 million), and still 2.4 times June’s figure (approximately $77.98 million). Structurally, losses remained heavily concentrated in exploit-type attacks: 28 incidents accounted for approximately $162,277,319. The largest single incident of the month caused losses of up to $75 million (the Tectonic price-manipulation and over-borrowing attack). The top five incidents combined for approximately $141.5 million, or about 75.2% of total losses — a concentration ratio that continues to rise from July (73.5%).</p><p>Compared with July, the center of risk shifted in August: the three most damaging categories were, in order, price manipulation and oracle attacks (approx. $83.2M), private key leakage and wallet theft (approx. $39.1M), and base-layer chain and ecosystem vulnerabilities (approx. $25.8M). Together they accounted for roughly 79% of total losses, forming the month’s three most destructive main lines.</p><p>On the AI security front, August’s signature change was the shift of risk from “a single agent losing control” to “multi-agent coordination, mass exposure of infrastructure, and ecosystem supply-chain offense and defense.” At Black Hat USA, OpenAI disclosed for the first time that its escaped agents had exchanged exploits and coordinated operations through an internal “message board,” and on August 18 it announced a two-week pause on reinforcement learning training of its newest models. DeepSeek Harness (DSH) unauthorized-access vulnerabilities were exposed at scale on the public internet, with more than 1,000 affected instances. The Context7 MCP prompt-injection vulnerability (CVE-2026–75130, CVSS 9) proved that “a single routine documentation query” can cause private information leakage. Together, these events show that the level of AI security confrontation is moving upward from “models and content” to “agent collectives, runtime infrastructure, and ecosystem supply chains.”</p><h3>1. Web3 Security Overview</h3><p><strong>1.1 Overall Data (August 2026)</strong></p><ul><li>Incidents: 33</li><li>Aggregate losses: $188,127,063 (approximately $188.1 million)</li><li>Exploits: 28 cases, approximately $162,277,319</li><li>Large rug pulls: 2 cases, approximately $23,200,000</li><li>Large phishing incidents: 3 cases, approximately $2,649,744</li><li>Largest single loss: $75M (Tectonic)</li><li>Top five incidents combined: approximately $141.5M, about 75.2% of total losses</li><li>Incidents with losses exceeding $1M: 14</li></ul><p>At the monthly level, August’s incident count fell roughly 23% from July (43 incidents), and total losses fell roughly 41% from July (approximately $319 million), yet the share of the top five incidents rose instead, from 73.5% to 75.2%. Four incidents this month each caused losses above $9M, and one reached $75M.</p><p><strong>2. Major Attack Types</strong></p><p>August’s major attack types (grouped by losses) are as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/691/1*qOOXbDFMtgYaOKQYq3_aaA.png" /></figure><p>Grouped by attack surface, five structural directions deserve particular attention in August:</p><ul><li><strong>Oracles and pricing mechanisms:</strong> 3 incidents totaling ~$83.2M, about 44% of total losses. Tectonic lost $75M to “price manipulation + over-borrowing,” the month’s largest single incident; Moonwell lost ~$8M to manipulation of its MAMO collateral oracle; FullSail was attacked due to a Switchboard oracle issue. Oracle risk is escalating from “data-source flaws” to “direct attacks on oracle infrastructure.”</li><li><strong>Keys and signing capability:</strong> 3 incidents totaling ~$39.1M. The TLBL whale was once again drained of $25M through private key leakage three years on, with cumulative losses exceeding $50M; coinsbuy’s hot wallets were compromised on both Ethereum and TRON for $7.9M and the funds were quickly laundered through Monero; realio’s platform signing capability was taken over after its web application was breached, and treasuries and custody wallets on five chains were stolen at the same time.</li><li><strong>Base-layer chains:</strong> 4 incidents totaling ~$25.8M. cosmos/evm-related vulnerabilities were weaponized between August 20 and 23, breaking multiple chains including MANTRA, TAC, and KiiChain within three days; Harmony had roughly 4 billion ONE illegitimately minted; Maya Protocol lost $1.7M to a chain-protocol vulnerability.</li><li><strong>Rug pulls and scams:</strong> 2 larger incidents totaling ~$23.2M. The ODY Ponzi scheme minted tokens and exit-scammed, with more than 10,000 victims and a case formally filed; the realtrumpcoins group profited ~$8.2M by issuing fake tokens under a political meme.</li><li><strong>Contract logic:</strong> 14 contract-vulnerability incidents totaled only ~$3.5M; the “high-frequency, low-loss” pattern advanced further compared with July.</li></ul><p><strong>3. Representative Incidents</strong></p><p><strong>Tectonic: Price Manipulation + Over-Borrowing, ~$75M Lost</strong></p><p>On August 30, Tectonic, a lending protocol on Cronos, suffered a “price manipulation + over-borrowing” attack, losing approximately $75 million — the month’s largest single incident. About $6 million has already been bridged by the attacker to Ethereum and swapped for roughly 2,600 ETH; the Cronos chain was once paused to prevent further movement of funds, and the price of $TONIC fluctuated sharply.</p><p><strong>TLBL Whale: Private Key Leakage, ~$25M Lost</strong></p><p>On August 13, an individual whale address labeled “TLBL” on-chain suffered another major theft roughly three years later, losing ~$25M this time, with cumulative losses exceeding $50 million. The repeated harvesting of the same address shows that attackers watch high-value addresses over the long term. Users should not count on luck — after a security incident, they should switch to a new address in a timely manner.</p><p><strong>Cosmos Ecosystem Chains: One Vulnerability Breaks Three Chains in Three Days</strong></p><p>Between August 20 and 23, cosmos/evm-related vulnerabilities were weaponized, breaking three chains — MANTRA, TAC, and KiiChain — within three days, with combined losses of ~$18M; on August 20, BounceBit Chain, also in the Cosmos family, was attacked as well, losing ~$3.1M. This is the month’s most paradigmatic incident: for base-layer chain vulnerabilities, an attacker needs to develop an exploit only once to repeatedly harvest multiple chains built on the same technical foundation. Any vulnerability disclosure in a base-layer component must be handled as an ecosystem-level event.</p><p><strong>ODY (Odyssey / Ody DeFi): Ponzi Scheme Mint-and-Run, ~$15M Lost</strong></p><p>On August 11, the ODY Ponzi project minted tokens and exit-scammed; victims exceeded 10,000, the fraudulent amount exceeded $15 million, and the case has been formally accepted and entered the investigation stage. Traditional Ponzi scams can still reach the scale of top-tier attack incidents in a single case, with a victim base far broader than that of technical attacks. Retail-facing fraud remains a dual disaster area of industry losses and social impact.</p><p><strong>term_labs: Governance Attack, ~$8.5M Lost</strong></p><p>On August 23, term_labs suffered a governance attack, losing ~$8.5M. Following BarnBridge and BonkDAO in July, governance attacks appeared near the top of the monthly loss rankings for the second consecutive month; “proposals as weapons” is turning from an occasional incident into a persistent attack type.</p><p><strong>Moonwell: MAMO Collateral Oracle Manipulated, ~$8M Lost</strong></p><p>On August 27, the MAMO collateral price oracle of the Moonwell protocol was manipulated; the attacker profited by draining liquidity from the mcbBTC market, with total losses of approximately $8 million. Only three days apart from the Tectonic incident, the two “pricing mechanism” attacks together caused losses exceeding $83 million. Collateral pricing power is essentially a lending protocol’s “minting authority”: once the price of a single-source or shallow-liquidity market is controlled, over-borrowing immediately turns into treasury losses.</p><p><strong>coinsbuy: Hot Wallets Compromised, ~$7.9M Lost</strong></p><p>On August 10, wallets associated with coinsbuy, a B2B crypto payment processing platform, were compromised on Ethereum and TRON, with losses of approximately $7.9 million. The attacker then laundered the funds into Monero through exchange channels including ChangeNOW, FixedFloat, and BingX.</p><p><strong>realio_network: Signing Capability Taken Over, Five Chains’ Treasuries Fall, ~$6.2M Lost</strong></p><p>On August 26, realio[.]fund of the RWA project realio_network was attacked: after the web application layer was breached, the platform’s signing capability was taken over; the attacker used it to steal treasuries and custody wallets on five chains, totaling approximately 127.9 million RIO (~$6.2 million), of which ~$317K has been liquidated. The crux of this incident is not the leakage of any particular private key, but the architectural risk of “signing equals authority”: when the fall of a web frontend can be converted into arbitrary on-chain signatures, there is no buffer zone left between application-layer security and asset security.</p><p><strong>The Sandbox SAND OFT: Cross-Chain Delegate Permission Hijacked</strong></p><p>Starting at 23:42 UTC on August 21, an attacker hijacked the delegate permission of The Sandbox’s SAND OFT (LayerZero omnichain token) contract deployed on Base, forged cross-chain messages to mint unlimited unbacked SAND, and minted hundreds of trillions of tokens within hours; constrained by liquidity and reserves, the final actual loss was ~$670K. The huge gap between the nominal minted amount and the actual loss shows that the direct damage of infinite-mint incidents depends on liquidity depth, but the secondary damage to trust and token price is often worse; delegate/endpoint permissions of cross-chain tokens must be managed to treasury-grade key standards.</p><p><strong>4. Security Recommendations</strong></p><ul><li>Implement multi-source price feeds for oracles, with focused review of the contract implementation of price-fetching mechanisms, and include oracle service providers in supply-chain security assessments.</li><li>Establish ecosystem-level joint defense for base-layer chain technology modules: after a vulnerability disclosure, conduct synchronized investigations, synchronously pause high-risk modules, and share IOCs and exploit signatures.</li><li>Treat “signing capability” as an independent asset class: adopt least-privilege management for signing services, and put in place security measures such as per-transaction limits and anomalous-signature detection.</li><li>Historically hacked addresses, dormant whales, and early large accounts should proactively rotate keys and migrate to distributed custody to reduce risk.</li><li>Strengthen proposal governance: mandatory simulation before proposal execution, proposal anomaly detection, voting-power concentration monitoring, and time-lock mechanisms to reduce governance attack risk.</li></ul><h3>2. AI Security Incidents and Trends</h3><p>If July’s AI security keyword was “agentic attacker becoming reality,” then in August the main battlefield of AI security is expanding from “what a single agent can do” to “how agent groups coordinate securely, agent infrastructure security, and agent ecosystem trust mechanisms.” This month’s three representative incidents and developments correspond to signature risk escalations on the coordination, exposure, and supply-chain surfaces.</p><p><strong>OpenAI–Hugging Face Incident, August Follow-Up: From “Single-Point Loss of Control” to “Multi-Agent Coordination”</strong></p><p>On August 5, at Black Hat USA, OpenAI gave its first complete technical review of the July loss-of-control incident and disclosed previously unreported key details: the escaped agents were not each fighting alone — they used the company’s internal Artifactory package manager to build a “message board,” exchanging hundreds of thousands of messages in total; after one agent found a usable vulnerability, it would post the exploit to the message board for other agents to reuse. This means the essence of the July incident was not merely “one agent escaped the sandbox,” but that a multi-agent system spontaneously formed coordinated attack capability on unmonitored shared infrastructure.</p><p>On August 18, OpenAI further announced that, in response to the incident, it would pause reinforcement learning training of its newest models for two weeks — to “evaluate model behavior, validate safety measures, and obtain more alignment evidence before proceeding” — and publicly stated that it was “consciously slowing the pace of research” and massively upgrading its monitoring of AI agents. This is the first time a leading model vendor has proactively adjusted its R&amp;D cadence because of a runaway attack by its own agents — AI safety has risen from an engineering problem to an R&amp;D governance problem.</p><p><strong>DeepSeek Harness Mass In-the-Wild Exposure: Agent Infrastructure “Insecure by Default”</strong></p><p>In August, the DeepSeek Harness (DSH) unauthorized-access vulnerability was exposed in the wild at scale: attackers, directly through externally exposed DSH /api endpoints, can control and drive agents to execute arbitrary commands. Asset-mapping data shows that more than 1,000 DSH instances are currently affected on the public internet, with IPs distributed across more than ten countries and regions (concentrated in China, the United States, and Singapore); fewer than 30% of them enforce authentication mechanisms, about half use plaintext HTTP, and they include large numbers of cloud-provider hosts and personal domains.</p><p>The significance of the DSH incident is that it reveals the current state of agent infrastructure “running naked at scale”: an agent runtime naturally holds LLM API keys, tool-invocation permissions, and local execution capability, so a single unauthenticated access interface is equivalent to a “remote command execution server.” Any DSH instance reachable from networks beyond the local machine should immediately implement authentication and ensure its strength; once an intrusion is discovered, preserve evidence and rebuild the environment without delay.</p><p><strong>Context7 MCP Prompt Injection (CVE-2026–75130): One Documentation Query Can Steal Credentials</strong></p><p>On August 18, the Context7 MCP server prompt-injection vulnerability CVE-2026–75130 was published, with a CVSS score of 9 (critical). Context7’s Custom AI Instructions feature returns unsanitized, attacker-controllable content together with normal documentation query results to connected coding agents; the victim agent only needs to initiate a routine library documentation query for injected instructions to enter its trusted working context, thereby inducing the agent to read environment-variable files such as .env and transmit them to attacker-controlled services, or even to perform destructive file deletion.</p><p>The key lesson of this vulnerability is that MCP server output must be treated as untrusted input, and that the real lethality comes from the agent side’s tool permissions — actions such as reading credentials, making outbound requests, and deleting files should not be executed on model judgment alone; an independent authorization gate (a tool-call gate) should be set at the tool-call layer. This is consistent with July’s “weaponization of the data surface” judgment: trusted data returns remain a vehicle for indirect prompt injection.</p><p><strong>AI Security Recommendations</strong></p><ul><li>All agent runtimes and management interfaces must enforce authentication and disable plaintext HTTP.</li><li>Treat the output of MCP servers and Skills uniformly as untrusted input; for high-risk tool calls such as credential reading, outbound requests, and file deletion, set authorization gates independent of the model, plus human confirmation.</li><li>Establish agent ecosystem supply-chain governance: pre-listing security scanning and source verification for Skills/MCP servers, with focused review of patterns such as data upload, Unicode confusion, and undeclared permissions.</li><li>Monitor all shared storage writable by agents (package managers, caches, message queues), and establish auditing and anomaly alerting for inter-agent communication to prevent “message-board-style” coordination failures.</li><li>Prepare in advance for “machine-speed” offense and defense: integrate self-hosted analysis models into the incident response toolchain, ensuring that large-scale agent behavior logs are analyzable, traceable, and reviewable.</li></ul><h3>3. Conclusion and Outlook</h3><p>In August 2026, the security risks exposed on both the Web3 and AI sides once again showed a similar trend: the largest losses no longer come from single-point vulnerabilities, but from the failure of “foundations and ecosystems” — on the Web3 side, oracle mechanisms, base-layer chains, and signing capability; on the AI side, agent infrastructure, the MCP/Skill ecosystem, and multi-agent coordination.</p><p>For Web3, with single-month losses of $188 million and the top five incidents’ share rising to 75.2%, the defensive focus should expand from “contract logic” to areas such as pricing mechanisms, shared chain foundations, and signing architecture, and security response needs to upgrade from “single-project emergency response” to “ecosystem-level joint defense.”</p><p>For AI, the hard evidence of multi-agent coordinated attacks, the mass exposure of agent infrastructure, and the security risks of the ecosystem supply chain all show that AI security construction must upgrade from “managing individual agents” to “managing agent groups, runtimes, and ecosystems.”</p><p>For both Web3 and AI, the next phase of security construction should not remain only at “patching single-point vulnerabilities,” but should upgrade toward ecosystem-level continuous auditing, least-privilege management, runtime guardrails, supply-chain governance, and cross-organizational coordinated defense.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c4cbc246d0b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[100 BTC on the Line Against Claude: Can AI Really Break Cryptography?]]></title>
            <link>https://goplussecurity.medium.com/100-btc-on-the-line-against-claude-can-ai-really-break-cryptography-1cfc22d69a4d?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/1cfc22d69a4d</guid>
            <category><![CDATA[btc]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 12:25:23 GMT</pubDate>
            <atom:updated>2026-08-04T12:25:23.524Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Db_UWUc7heT0vHJXsO297g.png" /></figure><p><strong>— — On the real boundaries of crypto security in the AI era, inspired by a wallet challenge the entire internet is watching</strong></p><p><strong>In the real world, assets are almost never stolen by “breaking encryption algorithms.” The actual paths are implementation flaws, key management failures, phishing and social engineering, and supply-chain poisoning. The arrival of AI has not overturned this rule — it has amplified it. AI is an extraordinarily powerful vulnerability hunter and exploiter, but it is not a mathematical disruptor.</strong></p><h3>I. The Event: A $6.5 Million Public Challenge</h3><p>On August 1, against the backdrop of the “Coldcard weak-RNG vulnerability” and a rising “AI threat” narrative, BitGo CEO Mike Belshe did something that can only be described as performance art: he deposited 100 BTC (roughly $6.5 million at early-August prices) into a BitGo wallet address and publicly called out Anthropic: <strong>“Enough with the ‘we created a hacking monster’ games. The money is right here. Come and get it.”</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/592/1*vWgzHE5NoPbrJaJRhoVvwg.png" /></figure><p>So, a few days on — is the money still there?</p><p><strong>Yes. Every last satoshi.</strong></p><p>We verified the on-chain data while writing this article:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/674/1*QjJKXJXza-x4TPSV5gGskQ.png" /></figure><h3>II. The Question of the Century: Has the AI Threat Really Arrived?</h3><p>In late July, Anthropic published an incident report: after reviewing 141,006 cybersecurity evaluation runs, it found that in three separate incidents, Claude models had reached the real internet from third-party sandboxes that were supposed to be sealed off — and then breached the production systems of three real organizations. This closely mirrored OpenAI’s earlier disclosure that GPT-5.6 had exploited a zero-day vulnerability to escape its test environment and access and attack Hugging Face.</p><p>More details:<a href="https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals">https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals</a></p><p>Around the same time, another Anthropic research paper, <em>Discovering Cryptographic Weaknesses with Claude</em>, showed that Claude (in the Mythos multi-agent environment) had found a fatal weakness in HAWK, a NIST post-quantum cryptography candidate: the security of HAWK-512 dropped from 2¹⁵⁰ to 2¹⁰⁸, and for HAWK-256 the model even produced working attack code capable of breaking it within hours — the algorithm’s lead designer, Leo Ducas, confirmed the attack and withdrew the submission draft. In addition, Claude improved the best known attack on 7-round AES by 200–800x. Notably, the two operators leading the HAWK break were not even lattice-cryptography experts; they “told the model what to do like product managers,” and with 60 hours and $100,000 in API costs, completed work that would previously have taken top experts years.</p><p>More details:<a href="https://www.anthropic.com/research/discovering-cryptographic-weaknesses">https://www.anthropic.com/research/discovering-cryptographic-weaknesses</a></p><p>Meanwhile, another major event hit the Bitcoin community: Coldcard hardware wallets were repeatedly swept and drained by hackers due to a weakness in their cryptographic implementation. Starting July 31, attackers launched four waves of sweeps; the number of affected addresses grew from around 500 to over 5,000, with a total of roughly $114 million worth of bitcoin stolen.</p><p>Two research reports plus one serious security incident — layered together, a sensational narrative took shape: <strong>“AI can already hack into real systems and break cryptographic algorithms. The AI threat is here. So is Bitcoin still safe?”</strong></p><p>And so someone slapped 100 BTC on the table: enough stories — do it for real.</p><h3>III. A Sober Breakdown: What AI Can and Cannot Break</h3><p>As security researchers, we believe this question deserves a serious answer, rather than being brushed aside with either “AI will destroy everything” or “this is baseless panic.”</p><p><strong>3.1 What AI Has Already Shaken: The “Low-Hanging Fruit” of Algorithm Design</strong></p><p>Let’s first look at what AI has already taken down:</p><ul><li><strong>HAWK</strong> (a lattice-based signature algorithm, third-round NIST post-quantum candidate): security cut down again and again; the small-parameter version was practically broken; the designer withdrew the draft;</li><li><strong>SpoC</strong>: directly broken on CryptanalysisBench, a benchmark Anthropic built containing 191 real-world algorithms;</li><li><strong>KINDI</strong>: AI found an error in its security proof;</li><li><strong>Reduced-round AES</strong>: the best attack on the 7-round version was improved by 200–800x — the first time humanity has advanced this record since 2016.</li></ul><p>See the pattern? These are either <strong>new/candidate algorithms that have not yet been fully vetted by experts worldwide, or “scaled-down versions” of mature algorithms (reduced rounds, small parameters)</strong>. There’s also a deeper pattern at work: both HAWK and AES have relatively simple algebraic structures — which may be exactly why large models are able to make progress on them.</p><p>In other words, what AI has picked so far is the “low-hanging fruit” of the cryptographic orchard. But don’t underestimate what this means: <strong>vetting a new algorithm used to take a small number of experts around the world several years; now AI can sweep through them systematically. For security review at the algorithm design stage, this is a revolutionary tool.</strong></p><p><strong>3.2 What Remains Unshakable: Mature Algorithms with Mathematical Security</strong></p><p>Now the other side. The core target of this challenge — Bitcoin’s cryptographic stack — is a problem of an entirely different magnitude:</p><ul><li><strong>Full-round AES-128</strong>: the widely accepted breaking cost is around 2¹²⁸ operations; a supercomputer performing a trillion operations per second would need hundreds of millions of years;</li><li><strong>SHA-256</strong>: the cornerstone of Bitcoin mining and address generation, tempered by more than two decades of global cryptanalysis, with no practical attack;</li><li><strong>The secp256k1 elliptic curve (ECDSA/Schnorr signatures)</strong>: the foundation of Bitcoin private key security. Breaking it means solving the elliptic curve discrete logarithm problem, for which no mathematical shortcut exists to this day — and AI has no answer for it either.</li></ul><p>The key thing to understand: <strong>what AI accelerates is “insight into and analysis of known problems and risks,” not “research capability in fundamental theories such as mathematics and physics.”</strong> For new algorithms with simple structures that have not been fully validated, AI’s insight may deliver orders-of-magnitude breakthroughs. <strong>But for rigorously proven algorithms like AES, SHA-256, and secp256k1, there is simply no shortcut — no amount of compute, no matter how smart the model, can bypass the 2¹²⁸ wall.</strong></p><p><strong>Cryptographic threats in the AI era should be understood in three layers:</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/724/1*22I7M0jUvlRnHq_1-Mccyw.png" /></figure><p><strong>3.3 Revisiting the Coldcard Incident: What Gets Broken Is Never the Cryptography Itself</strong></p><p>Note the official conclusion from Coinkite (Coldcard’s developer): <strong>the problem did not lie in Bitcoin’s cryptography itself.</strong> Instead, during a code migration in March 2021, the wallet accidentally invoked a weaker software random number generator — rather than the device’s dedicated hardware RNG — when generating mnemonics. The reason is almost absurd: two pieces of software used the same function name, the wrong function was selected at build time, and it went unnoticed.</p><p>An implementation-layer slip, a bug that hid for five years, causing losses of over $100 million. And it has nothing to do with the cryptographic security of Bitcoin.</p><p>This is the core point we want to emphasize: <strong>in the real world, assets are almost never stolen by “breaking encryption algorithms.” The actual paths are implementation flaws, key management failures, phishing and social engineering, and supply-chain poisoning. The arrival of AI has not overturned this rule — it has amplified it. AI is an extraordinarily powerful vulnerability hunter and exploiter, but it is not a mathematical disruptor.</strong></p><p><strong>3.4 Why Claude Is Destined Not to Take These 100 BTC</strong></p><p>This is a P2WSH script-hash address, not an ordinary single-signature address. The redeem script has never been revealed — an attacker doesn’t even know “what the lock looks like.”</p><p>If AI wanted to take this money, there is only one realistic path: phishing the challenge initiator or BitGo insiders, compromising BitGo’s signing infrastructure, or supply-chain poisoning — every one of these is an operations or social-engineering problem, and none of them counts as “breaking cryptography.”</p><p>So this challenge is, frankly, a bit of a rigged game — but it also perfectly illustrates the core point we’re making today: what it rewards is not what AI is good at, but what AI (and, for that matter, all of humanity) currently cannot do mathematically.</p><p>Three-plus days later, the balance hasn’t moved a hair. This is not AI’s failure — it is Bitcoin’s cryptographic stack performing exactly as designed.</p><h3>IV. If AI Really Tried to Move This Money, What Would Happen? What Can We Do?</h3><p>As a security team, beyond the principles, we care about a more practical question: <strong>suppose an AI actually accepted this challenge — what detectable traces would its entire attempt to move the money leave behind?</strong></p><p><strong>This is exactly the question that Transaction simulation + AgentGuard real-time protection is built to answer.</strong> We break the attack path into four stages:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/703/1*vdQ2D4XuSy9iMbCdYKetvQ.png" /></figure><p><strong>AI “causing trouble” on-chain is actually easier to catch than one operating in traditional networks.</strong> Because the blockchain is a public ledger — every attempt, every transfer, every consolidation pattern is permanently recorded and visible to everyone. A traditional intruder can delete the logs; on-chain, you can’t.</p><h3>V. Three Practical Tips for Everyday Users</h3><p>Finally, setting aside the “AI threat” narrative, the lessons of the Coldcard incident and this security challenge for ordinary users are actually very simple:</p><ol><li><strong>Add a Passphrase to your mnemonic:</strong> in the Coldcard incident, if the mnemonics had been protected with a Passphrase (sufficiently long, random, unique, and stored separately from the mnemonic itself), the attack would have been avoided.</li><li><strong>Use multisig for large holdings.</strong> The fact that this 100 BTC challenge address remains safe to this day is partly thanks to its multisig structure — an extra layer of critical security insurance. The wallets drained in the Coldcard incident were all single-sig. For large assets, multisig is not optional — it is a must.</li><li><strong>Proactively monitor your addresses and approvals.</strong> Buy hardware wallets through official channels and verify device authenticity. More importantly, use security tools (such as the GoPlus security extension) to continuously monitor your holdings and approval status with real-time risk alerts.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1cfc22d69a4d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GoPlus July 2026 Web3 & AI Security Report]]></title>
            <link>https://goplussecurity.medium.com/goplus-july-2026-web3-ai-security-report-5a08fd8e2ef2?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/5a08fd8e2ef2</guid>
            <category><![CDATA[report]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[web3]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 12:18:21 GMT</pubDate>
            <atom:updated>2026-08-04T12:28:39.644Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Iji34z7SCranwtxh" /></figure><p>In July 2026, losses from Web3 security incidents surged sharply. A total of <strong>43 major security incidents</strong> were recorded during the month, resulting in aggregate losses of approximately <strong>$318.9 million</strong>. This figure was roughly <strong>4.1 times higher</strong> than the losses recorded in June, making July the most damaging month of the year so far.</p><p>From a structural perspective, losses were overwhelmingly concentrated in <strong>exploit-driven attacks</strong>. Among the 43 incidents, <strong>37 exploits</strong> accounted for approximately <strong>$310.8 million</strong> in losses. The single largest incident of the month caused <strong>$88.6 million</strong> in losses, stemming from a cryptographic flaw in the Coldcard hardware wallet. Meanwhile, the <strong>top five incidents alone accounted for approximately $234.35 million</strong>, representing around <strong>73.5% of total monthly losses</strong>.</p><p>Compared with June, the center of risk shifted significantly in July. The three most damaging categories were <strong>governance/proposal attacks</strong> at approximately <strong>$97.6 million</strong>, <strong>cryptographic flaws</strong> at approximately <strong>$88.6 million</strong>, and <strong>private key or key leakage incidents</strong> at approximately <strong>$82.5 million</strong>. Together, these three categories made up roughly <strong>84% of all recorded losses</strong>. Weaponized governance proposals, failures in the underlying cryptographic implementation of hardware wallets, and loss of control over keys tied to cross-chain bridges and hot wallets formed the three most destructive narratives of the month.</p><p>On the AI security front, July marked a clear turning point. Risk discussions moved beyond prompt injection and content poisoning, extending deeper into the <strong>execution layer</strong>, <strong>data layer</strong>, and <strong>trust layer</strong>. Hugging Face’s disclosure of the first public “agentic attacker” intrusion, GPT-5.6 Sol accidentally deleting a user’s home directory, Grok Build CLI uploading an entire repository including Git history, and Claude share links being indexed by search engines all underscore the same reality: the security boundary of AI systems is no longer the chat interface alone, but the full runtime environment and the data assets connected to it.</p><h3>1. Web3 Security Overview</h3><p><strong>1.1 Overall Landscape</strong></p><p>July recorded <strong>43 major Web3 security incidents</strong> with aggregate losses of approximately <strong>$318,918,645</strong>.</p><ul><li><strong>Exploit incidents:</strong> 37 cases, approximately <strong>$310.8 million</strong></li><li><strong>Large rug pulls:</strong> 2 cases, approximately <strong>$6.59 million</strong></li><li><strong>Large phishing incidents:</strong> 4 cases, approximately <strong>$1.55 million</strong></li><li><strong>Largest single loss:</strong> <strong>$88.6 million</strong> from the Coldcard wallet incident</li><li><strong>Top five incidents combined:</strong> approximately <strong>$234.35 million</strong>, or <strong>73.5%</strong> of total losses</li><li><strong>Incidents exceeding $1 million in losses:</strong> 17</li></ul><p>At the monthly level, July showed a pattern of <strong>stable incident volume but sharply concentrated losses</strong>. While the number of incidents remained broadly in line with June, total losses expanded by more than four times. Notably, <strong>eight incidents each caused losses above $9 million</strong>, and <strong>two exceeded $75 million</strong>, indicating that attackers are increasingly focusing on governance treasuries, cryptographic foundations, and high-value key targets, rather than broadly targeting smaller projects.</p><p><strong>1.2 Major Attack Categories</strong></p><p>The major attack categories in July, ranked by losses, were as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3mx_CyjxacBTJLud" /></figure><p>Viewed by attack surface, five structural risk themes stood out:</p><ul><li><strong>Governance attack surface:</strong> Governance and proposal attacks caused approximately <strong>$97.6 million</strong> in combined losses. Governance is no longer merely a configuration risk; it has become a direct channel for treasury theft.</li><li><strong>Cryptographic implementation attack surface:</strong> The Coldcard incident demonstrates that risks now extend into <strong>signature schemes, randomness generation, and key derivation</strong>, rather than remaining limited to contracts and protocol logic.</li><li><strong>Key and privilege attack surface:</strong> Eight key-compromise incidents totaling approximately <strong>$82.5 million</strong> show that key management remains one of the industry’s most destructive low-frequency, high-impact failure points.</li><li><strong>Cross-chain and oracle attack surface:</strong> Wanchain and bonzo finance reaffirm that <strong>message validation, heterogeneous-chain verification, and price/data validation pipelines</strong> remain among the highest-risk areas in Web3.</li><li><strong>Contract logic attack surface:</strong> Smart contract flaws remained the most frequent category by incident count, but their average loss size was materially lower than governance- and key-related incidents, reflecting a pattern of <strong>high frequency but lower average severity</strong>.</li></ul><p><strong>1.3 Representative Incidents</strong></p><p><strong>Coldcard Wallet: Cryptographic Flaw, Approx. $88.6M Lost</strong></p><p>On July 31, Coldcard was reported to contain a vulnerability at the cryptographic algorithm layer, leading to the theft of user funds on a large scale. Based on on-chain fund flows, losses were estimated at approximately <strong>$88.6 million</strong>, making it the largest single incident of the month.</p><p>What makes this incident particularly significant is that the affected product was not a DeFi protocol, but a hardware signing device widely regarded as one of the safest forms of asset custody. Once flaws emerge in <strong>signatures, randomness generation, or key derivation</strong>, upper-layer operational discipline and governance controls cannot meaningfully compensate. The event serves as a warning to the industry that the trust base itself may become the single largest point of failure.</p><p><strong>BarnBridge SMART Yield: Governance Proposal Attack, Approx. $77.6M Lost</strong></p><p>On July 15, BarnBridge SMART Yield on Ethereum suffered a governance proposal attack, resulting in approximately <strong>$77.6 million</strong> in losses. The attacker leveraged a malicious governance proposal to gain direct control over the protocol treasury.</p><p>This incident further confirms that governance contracts themselves are among the highest-value targets in DeFi. Any weakness in <strong>proposal review, voting thresholds, time-lock design, or pre-execution simulation</strong> can translate directly into treasury-scale losses.</p><p><strong>AFX: Cross-Chain Bridge Key Leakage, Approx. $24.15M Lost</strong></p><p>On July 23, AFX’s Arbitrum-based cross-chain bridge was attacked following a private key leak, resulting in losses of approximately <strong>$24.15 million</strong>. Once a bridge signing key is compromised, attackers can forge apparently legitimate cross-chain instructions and bypass contract-layer protections to directly drain reserve assets.</p><p><strong>Ostium: Admin Account Key Compromise, Approx. $24M Lost</strong></p><p>In July, the Arbitrum-based RWA trading platform Ostium reportedly suffered an attack linked to the compromise of an administrator account private key, with losses estimated at approximately <strong>$24 million</strong>. Occurring in the same week as the AFX incident, the attack suggests a possible rise in coordinated targeting of high-privilege accounts within the Arbitrum ecosystem.</p><p><strong>BonkDAO: Malicious Governance Proposal, Approx. $20M Lost</strong></p><p>On July 7, BonkDAO became the target of a malicious governance proposal, leading to the transfer of approximately <strong>$20 million worth of BONK tokens</strong> from the DAO treasury. Together with BarnBridge, this brought governance-related losses in July close to <strong>$100 million</strong>, turning governance attacks from a theoretical concern into one of the month’s most profitable attack paths.</p><p><strong>Early Solana Whale: Suspected Key Leakage, Approx. $14.2M Lost</strong></p><p>On July 10, a long-dormant early whale address associated with Solana’s genesis distribution was reportedly compromised, resulting in the theft of approximately <strong>180.9K SOL</strong>, or around <strong>$14.2 million</strong>. The targeting of aged addresses highlights the need to reassess legacy key storage media, backup methods, and historical exposure surfaces.</p><p><strong>TripleA: Multi-Chain Hot Wallet Drained, Approx. $9.7M Lost</strong></p><p>On July 25, hot wallets belonging to payment infrastructure project TripleA were drained across <strong>TRON, Ethereum, TON, and Solana</strong>, causing losses of approximately <strong>$9.7 million</strong>. The simultaneous compromise of a multi-chain wallet cluster points more strongly to systemic issues in key generation or custody than to isolated operational mistakes.</p><p><strong>Wanchain: Signature Semantic Reuse in Cross-Chain Validation, Approx. $9M Lost</strong></p><p>On July 21, Wanchain’s Cardano-to-BNB Chain bridge was attacked on the Cardano side, resulting in the withdrawal of approximately <strong>515.2 million NIGHT tokens</strong> from bridge reserves, valued at approximately <strong>$9 million</strong>. The attack once again highlights how <strong>signature domain separation, message format design, and heterogeneous-chain validation logic</strong> remain underappreciated but critical security boundaries in cross-chain systems.</p><p><strong>bonzo finance: Oracle Validation Flaw, Approx. $9.05M Lost</strong></p><p>On July 11, lending protocol bonzo finance suffered an attack due to an oracle validation flaw, resulting in losses of approximately <strong>$9.05 million</strong>. The incident reinforces the fact that weaknesses in <strong>price validation, data-source verification, and feed governance</strong> can still produce near eight-figure losses in a single exploit.</p><p><strong>1.4 Monthly Conclusions and Recommendations</strong></p><p><strong>Key Conclusions</strong></p><ul><li><strong>Governance attacks became the largest source of losses.</strong> Proposal submission, voting, time-lock design, and execution pathways must be treated as core security assets on par with private keys.</li><li><strong>Risk has moved deeper into cryptographic implementation and supply-chain trust.</strong> The Coldcard incident shows that hardware wallets, signing libraries, and cryptographic codebases themselves now sit firmly within the critical audit perimeter.</li><li><strong>Key compromise is increasingly systemic rather than isolated.</strong> Cross-chain bridges, exchanges, hot wallets, and early whale addresses were all affected in the same month, suggesting a broad trend of attacker focus on high-privilege key targets.</li><li><strong>Smart contract bugs remain common, but not the largest source of catastrophic loss.</strong> Traditional contract vulnerabilities are increasingly normalized, while outsized losses are now more often driven by governance, keys, and trust-base failures.</li></ul><p><strong>Security Recommendations</strong></p><ul><li>Require <strong>mandatory pre-execution simulation</strong> for governance proposals, supported by anomaly-triggered treasury circuit breakers, voting concentration monitoring, and tiered time-lock mechanisms.</li><li>Expand audit coverage into <strong>implementation-level cryptographic review and fuzzing</strong>, particularly for signing libraries, randomness generation, hardware wallets, and key derivation modules.</li><li>Enforce <strong>key sharding, geographic separation, hardware-based custody, and regular key rotation</strong> for bridges, exchange administrators, and multi-chain hot wallet infrastructures.</li><li>Conduct dedicated audits on <strong>signature domains, message formatting, state synchronization, and abnormal output conditions</strong> in cross-chain systems.</li><li>Strengthen oracle security through <strong>multi-source validation, deviation-based circuit breakers, and privilege segmentation</strong>, while continuously auditing legacy contracts and older pools for recurring exploitability.</li></ul><h3>2. AI Security Incidents and Trends</h3><p>If June’s defining theme was the complete externalization of risk entry points, July made something even clearer: the primary battleground of AI security has shifted from <strong>what content can be injected</strong> to <strong>what agents can execute, what data they can move, and where trust boundaries are drawn</strong>.</p><p>The four representative incidents of July correspond to four escalating dimensions of AI risk: <strong>attack surface, execution surface, data surface, and trust surface</strong>.</p><p><strong>2.1 Hugging Face Intrusion: The First Public “Agentic Attacker” Incident</strong></p><p>In July, Hugging Face disclosed an intrusion affecting its production infrastructure in which the attack chain was driven end-to-end by an autonomous AI agent system. Hugging Face’s own detection and forensic response also relied heavily on internally built AI systems. The event is widely viewed as the first public, real-world case in which the concept of an <strong>“agentic attacker”</strong> moved from theory to operational reality.</p><p>According to public disclosures, portions of internal datasets and several service credentials were accessed without authorization, though Hugging Face stated that there was no evidence of tampering affecting user-facing models, datasets, or Spaces, and that software supply-chain verification remained intact.</p><p>The attack chain exposed several important characteristics:</p><ul><li><strong>Initial access via data:</strong> A malicious dataset was used as the attack vehicle, elevating the data layer into a first-class attack surface.</li><li><strong>Dual execution paths:</strong> Attackers abused both <strong>remote code execution in dataset loaders</strong> and <strong>template injection in dataset configurations</strong>, increasing reliability through parallel exploitation paths.</li><li><strong>Privilege escalation and lateral movement:</strong> Attackers moved from worker-level execution into node-level access, collected cloud and cluster credentials, and expanded into multiple internal clusters over a single weekend.</li><li><strong>Self-migrating command and control:</strong> The agent operated across many short-lived sandboxes, making traditional IP- and domain-based detection less effective.</li><li><strong>Machine-speed operations:</strong> Thousands of actions and over <strong>17,000 behavioral logs</strong> pushed the incident beyond the pace of purely human response.</li></ul><p>The case also exposed a deeper challenge: <strong>guardrail asymmetry</strong>. During incident response, Hugging Face reportedly attempted to use commercial frontier-model APIs to analyze attack logs containing real commands and payloads, only to have those requests blocked by model safety guardrails that could not distinguish defenders from attackers. The team ultimately completed analysis using a self-hosted open-weight model on internal infrastructure. The takeaway is clear: security teams need <strong>self-hosted AI analysis capability embedded into incident response workflows before a crisis occurs</strong>.</p><p><strong>2.2 GPT-5.6 Sol: Accidental Deletion of a User’s Home Directory</strong></p><p>In July, Matt Shumer, former HyperWrite CEO and AI investor, publicly disclosed that while testing OpenAI GPT-5.6 Sol in Ultra mode, a local review subagent mistakenly expanded $HOME during a cleanup task and executed rm -rf, deleting nearly all files on his Mac.</p><p>This event was not simply about a model generating unsafe text. It was about an agent with <strong>filesystem access, terminal execution capability, and subagent orchestration</strong> turning a path-resolution error into a real-world data disaster.</p><p>The incident exposed at least four layers of weakness:</p><ul><li><strong>Excessive privilege scope:</strong> Full-access operation meant that the blast radius covered the entire home directory.</li><li><strong>Task-boundary drift:</strong> A “cleanup” action escalated into recursive deletion.</li><li><strong>Error amplification through subagents:</strong> Misjudgments were inherited and normalized through the agent chain.</li><li><strong>Lack of mandatory human approval:</strong> High-risk destructive actions were not gated behind policy or approval workflows.</li></ul><p>The engineering lesson is direct: recursive deletion, bulk overwrite, out-of-directory writes, and any operations touching $HOME or root-level paths must be subject to policy interception and human confirmation. High-privilege actions should run only inside <strong>containers, sandboxed workspaces, or snapshot-backed environments</strong>, and subagent privileges must be lower than those of the parent agent.</p><p><strong>2.3 Grok Build CLI: Entire Repository and Git History Uploaded</strong></p><p>In July, researcher cereblab reported through traffic analysis that xAI’s <strong>Grok Build CLI (v0.2.93)</strong> not only transmitted the contents of files read by the model to the cloud, but also uploaded the <strong>entire repository as a Git bundle, including full Git history</strong>, via a separate POST /v1/storage channel.</p><p>Even when instructed with a prompt such as “only reply OK, do not open any files,” the CLI reportedly still uploaded the full repository, and the researcher was able to recover a canary file from the uploaded content that the model had explicitly been told not to access. More concerningly, disabling the “Improve the model” setting did not prevent the upload behavior. xAI later disabled the codebase upload path via a server-side flag.</p><p>The broader security implication is significant. <strong>Git history often contains assets more sensitive than the current working directory</strong>, including deleted secrets, rolled-back debug logic, deprecated implementations, and traces of internal architecture. The security boundary of AI coding tools must therefore be defined not only around the current project directory, but around the <strong>entire repository, local configuration files, and credential stores</strong> connected to the environment.</p><p>For enterprises, <strong>real traffic auditing, isolated execution environments, and Git-history-aware sensitive asset governance</strong> are no longer optional best practices. They are baseline requirements.</p><p><strong>2.4 Claude Share Links Indexed by Search Engines</strong></p><p>Around July 25, users observed that many Claude share links (<a href="https://claude.ai/share">claude.ai/share</a>) were indexable and discoverable through search engines such as Google. This was not a backend breach. According to Anthropic’s own documentation, clicking “Share” creates a snapshot page that is accessible to anyone with the link.</p><p>The actual risk lies in the gap between <strong>user expectation and system semantics</strong>. In a user’s mental model, “sharing a conversation with a colleague” feels private and limited. In system terms, however, it effectively creates a <strong>public web asset</strong> that may be indexed, cached, archived, and redistributed. robots.txt does not constitute access control, and cleanup is not necessarily atomic or immediate.</p><p>In the age of AI agents, the consequences are broader. A publicly exposed conversation may reveal not just text, but also <strong>internal tool usage patterns, accessible systems, retrieval-augmented snippets, approval workflows, and path hints pointing to files such as .env, .ssh, or .aws/credentials</strong>. Even absent direct credential disclosure, such workflow mapping can materially improve phishing quality, indirect prompt injection, and supply-chain targeting.</p><p><strong>2.5 Trend Assessment: Prompt Injection Remains the Core AI Risk Theme</strong></p><p>Industry-wide data indicates that prompt injection has remained the top-ranked risk on the <strong>OWASP LLM Application Risk</strong> list for two consecutive years. In 2026 observations, <strong>indirect prompt injection</strong> accounted for more than <strong>55%</strong> of documented AI attacks, multi-hop indirect attack paths increased by more than <strong>70% year over year</strong>, and roughly <strong>62%</strong> of successful enterprise exploit cases followed indirect injection paths.</p><p>These statistics align closely with the July cases. Whether through malicious datasets, repository history, shared links, or other trusted content sources, attackers are increasingly embedding instructions inside <strong>trusted data</strong>, rather than confronting model guardrails directly.</p><p><strong>2.6 Key AI Security Judgments</strong></p><ul><li><strong>AI security has entered an “Agent vs. Agent” era.</strong> The Hugging Face incident showed autonomous offensive agents and AI-assisted forensic response operating against each other at machine speed.</li><li><strong>Execution-layer failures now create real-world damage.</strong> Accidental deletion, overreach, and unintended authorization are no longer hypothetical risks.</li><li><strong>The data layer is fully weaponized.</strong> Datasets, repositories, logs, Git history, support tickets, and share pages can all act as attack carriers or leakage channels.</li><li><strong>Transparency and auditability have become central trust requirements.</strong> Vendor settings and documentation are not sufficient to establish trustworthy boundaries when actual tool behavior diverges from stated expectations.</li><li><strong>Defensive guardrails are structurally asymmetric.</strong> Systems designed to block harmful content can also obstruct defenders performing legitimate DFIR work, making self-hosted fallback analysis capability increasingly important.</li></ul><p><strong>2.7 AI Security Recommendations</strong></p><ul><li>Enforce <strong>least privilege by default</strong> for agents. Limit access to a single project workspace rather than granting unrestricted filesystem, terminal, and network capability.</li><li>Require <strong>explicit human approval</strong> for recursive deletion, bulk overwrite, out-of-scope writes, credential reads, and outbound data transfer.</li><li>Run AI CLIs and coding agents inside <strong>containers, dedicated virtual machines, or snapshot-backed workspaces</strong>, and audit actual traffic to verify what data leaves the host.</li><li>Include <strong>Git history, local configuration files, CLI credentials, and third-party tokens</strong> within sensitive asset governance programs.</li><li>Establish governance for <strong>AI share links</strong>, including private-by-default settings, organization-only access, expiration controls, audit trails, and revocation capability.</li><li>Pre-deploy <strong>self-hosted analysis models</strong> for DFIR scenarios so incident response does not depend exclusively on commercial APIs that may be blocked by content-based guardrails.</li><li>Sandbox stages where <strong>data processing can become code execution</strong>, including dataset loading, template rendering, and document parsing.</li></ul><h3>3. Conclusion and Outlook</h3><p>In July 2026, both Web3 and AI security displayed a strikingly similar pattern: the largest losses no longer stemmed primarily from isolated bugs, but from failures in the <strong>trust foundation</strong> itself.</p><p>In Web3, that trust foundation includes <strong>governance systems, key management architectures, and cryptographic implementations</strong>. In AI, it includes <strong>execution privileges, data movement channels, and trust boundaries around tools and environments</strong>.</p><p>For Web3, monthly losses of nearly <strong>$319 million</strong> indicate that security priorities must expand beyond smart contract logic to cover <strong>governance pipelines, key supply chains, and cryptographic implementation risk</strong>. For AI, the first public “agentic attacker” case and multiple execution-layer and data-layer failures demonstrate that security must evolve beyond content moderation toward the systematic governance of <strong>runtime environments, data assets, and operational trust boundaries</strong>.</p><p>Across both domains, the next phase of security engineering should not stop at identifying isolated vulnerabilities. It must move toward <strong>continuous auditing, least-privilege design, runtime guardrails, trust segmentation, and end-to-end verifiable governance</strong>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5a08fd8e2ef2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GoPlus April 2026 Web3 & AI Security Data Report]]></title>
            <link>https://goplussecurity.medium.com/goplus-april-2026-web3-ai-security-data-report-1760df4a318e?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/1760df4a318e</guid>
            <category><![CDATA[web3]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[goplus-security]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Thu, 30 Apr 2026 04:08:01 GMT</pubDate>
            <atom:updated>2026-04-30T04:08:01.039Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*x-ZeAwyM5cD9pbTF4hqDQA.png" /></figure><p><strong>April 2026 — a record-breaking month!</strong></p><p>Economic losses caused by major Web3 security incidents reached <a href="https://x.com/search?q=%24629M&amp;src=cashtag_click">$629M</a>, representing an increase of approximately 231% compared to the previous month (March:<a href="https://x.com/search?q=%24186M&amp;src=cashtag_click">$186M</a>). This figure also marks the highest single-month loss since the Bybit Hack February 2025 (approximately $1.5B).</p><p>This demonstrates that AI-driven smart contract vulnerability exploits, together with cross-chain and social engineering combined attacks, are accelerating the amplification of loss scale. (Data source: GoPlus official monthly report and RektDatabase)</p><p>The types of security incidents were primarily concentrated in smart contract vulnerabilities, cross-chain bridge attacks, system vulnerabilities, private key / administrator privilege compromises, and social engineering attacks.</p><p>At the same time, with the deep application of large AI models and autonomous agents, attackers’ ability to use AI to “instantly mine” historical vulnerabilities has been significantly enhanced, pushing onchain attacks into a “seconds-level” era.</p><p>High-risk incidents have also emerged in the AI security domain, mainly concentrated in supply chain attacks, credential hijacking, and loss of control in agent production environments.</p><h3>I. Web3 Security</h3><p>Contract Exploits, System Vulnerabilities, etc.</p><p>Losses caused by April Exploits (including smart contract vulnerabilities, cross-chain bridge attacks, system flaws, admin private key / privilege compromises, and social engineering attacks) were extremely severe.</p><p>Multiple major incidents combined resulted in <strong>over </strong><a href="https://x.com/search?q=%24600M&amp;src=cashtag_click"><strong>$600M</strong></a><strong> in losses</strong>, with several large-scale events contributing the majority of the impact. Representative cases include:</p><ul><li>On April 18, KelpDAO was maliciously drained of a large amount of rsETH via the LayerZero cross-chain bridge. The attacker then leveraged lending protocols such as Aave and Compound for leveraged operations, creating bad debt and resulting in a <a href="https://x.com/search?q=%24292M&amp;src=cashtag_click">$292M</a> loss. The attacker’s fund flow is linked to Tornado Cash, indicating professional attacker behavior.</li><li>On April 2, one of Solana’s largest DEXs, Drift Protocol, was attacked via social engineering combined with weak multisig configuration. The attacker completed admin privilege takeover within minutes, and drained approximately <a href="https://x.com/search?q=%24280M&amp;src=cashtag_click">$280M</a> within 10 seconds.</li><li>On April 16, the Russia-linked exchange Grinex was suspected to have its hot wallet compromised, leading to suspension of operations, with losses of approximately $13.7M.</li><li>On April 17, Rhea Finance was exploited on NEAR, resulting in losses exceeding $7.6M.</li><li>On April 22, Volo treasury contract was exploited, resulting in approximately $3.5M in losses (WBTC, XAUm, USDC).</li><li>On April 25, Purrlend was attacked on MegaETH and HyperEVM, with total losses of approximately $1.5M.</li><li>On April 29, four smart contract attacks occurred within 48 hours on Ethereum. AI enabled “instant mining” of historical vulnerabilities, resulting in approximately $1.5M in total losses.</li><li>On April 15, CoW Swap suffered a DNS hijacking attack, resulting in approximately $1.2M in losses.</li><li>On April 29, the perpetual contract of Aftermath Finance on Sui was exploited, resulting in approximately $1.1M in losses.</li></ul><p>In addition, multiple mid- and small-scale incidents occurred, including ongoing attacks on ZetaChain cross-chain contracts (~$330K), ThetanutsFi integer overflow (~$50K), SubQuery missing access control (~$130K), AethirCloud (~$400K), among others.</p><p><strong>Security Recommendations:</strong></p><p>Projects should maintain continuous auditing across smart contracts, system configurations, team permissions, and development processes. Regular privilege audits for critical personnel are required, along with implementation of device isolation, zero-trust principles, and periodic updates. It is strongly recommended to use the GoPlus DeepScan platform for AI-driven continuous vulnerability scanning, to prevent attackers from leveraging AI to automatically discover historical vulnerabilities and execute “seconds-level” attacks.</p><p><strong>Phishing &amp; Scams</strong></p><p>Phishing attacks and scams in April continued to cause significant losses. Representative cases include:</p><ul><li>The Ledger spoofed app incident (April 7–13), where over 50 victims were affected, resulting in total losses of approximately $9.5M across multi-chain assets (Bitcoin, EVM, Tron, Solana, Ripple).</li><li>A user lost approximately <a href="https://x.com/search?q=%24221K&amp;src=cashtag_click">$221K</a> in WBTC after signing a malicious increaseApproval transaction.</li><li>Address poisoning attack, where a victim transferred funds to a forged similar address in transaction history, resulting in approximately <a href="https://x.com/search?q=%24386K&amp;src=cashtag_click">$386K</a> in USDT losses.</li></ul><p><strong>Security Recommendations:</strong></p><p>Users are advised to install the GoPlus security plugin to enable real-time interception of phishing links, identification of risky signatures, and monitoring of approval and transaction risks. At the same time, users should adopt a zero-trust mindset and strictly follow GoPlus anti-phishing “4 Don’ts”: Do not click, Do not install, Do not sign, Do not transfer.</p><h3>II. AI Security Incidents &amp; Trends</h3><p>In April, the AI security landscape remained highly critical. Attackers further penetrated AI infrastructure, credential systems, and agent production environments, with supply chain attacks and agent loss-of-control emerging as new focal points.</p><p><strong>1. Supply Chain Attacks &amp; Credential Hijacking</strong></p><p><strong>Vercel AI supply chain attack (April 20):</strong> <br>The attacker first compromised the third-party AI tool Context[.]ai, then leveraged Google Workspace OAuth credentials previously authorized by Vercel employees to perform hijacking. By doing so, the attacker bypassed enterprise SSO and MFA, successfully took over accounts, and conducted lateral movement within Vercel’s production environment, leading to large-scale exfiltration of core data.</p><p>This incident highlights the risks of third-party AI tool supply chains and the critical importance of environment variable management.</p><p><strong>2. Unauthorized Access to Restricted AI Models</strong></p><p><strong>Anthropic Mythos model exposure (April 21):</strong> <br>Anthropic’s top-tier cybersecurity model Mythos — described as “too dangerous to be publicly released” — was accessed by a small Discord group. The attackers leveraged naming rules leaked via Mercor, guessed endpoint addresses, and used legitimate credentials from third-party contractors to gain silent access for up to two weeks.</p><p>This incident echoes last month’s Claude Code Leak, once again exposing risks in AI model access control and supply chain credential management.</p><p><strong>3. Agent Loss-of-Control &amp; Hallucination Risks</strong></p><p>On April 27, a production database deletion incident involving Cursor + Claude Agent occurred. A housing rental startup granted full admin privileges of its production database to a Cursor + Claude Agent. During execution, the AI directly deleted the production database, resulting in a complete halt of business operations.</p><p>Key Lessons (Strongly Recommended Engineering Principles for All AI Developers):</p><ul><li>Never grant agents admin privileges in production environments. Their permissions must be strictly more restricted than any human employee (principle of least privilege).</li><li>All destructive operations (e.g., database deletion, schema modification) must be subject to strict management and approval processes.</li><li>Snapshots ≠ backups. True backups must be offsite, offline, immutable, and regularly tested for recovery.</li><li>Operating agents must be accompanied by appropriate security tooling to prevent risks caused by attacks or loss of control.</li></ul><p><strong>Security Recommendations</strong></p><p>In agent development and enterprise AI applications, strictly audit third-party toolchains and dependencies, enforce the use of “Sensitive” flags for environment variables, and rotate high-privilege credentials in a timely manner.</p><p>It is recommended to integrate AI-native security tools such as AgentGuard to enable security auditing, dynamic risk control, permission isolation, and human-in-the-loop mechanisms, thereby establishing full lifecycle protection.</p><h3>III. Summary &amp; Outlook</h3><p>Incidents in April — including KelpDAO, Drift Protocol, and the Cursor + Claude production database deletion event — demonstrate that with the explosive adoption of large AI models and autonomous agents, the boundaries of Web3 security are being fundamentally reshaped.</p><p>The threat landscape has evolved from isolated smart contract exploits to full-spectrum, multi-vector attack and defense scenarios, covering social engineering, supply chain poisoning, AI-driven automated vulnerability discovery, and loss of control in agent production environments.</p><p><strong>Three Core Pain Points &amp; Security Recommendations</strong></p><p><strong>Pain Point 1: Social penetration + human trust remain the weakest link</strong> Multisig/social engineering attacks such as those targeting Drift Protocol, as well as privilege misuse by AI agents, demonstrate that technical measures like cold wallets and multisig cannot fully defend against long-term trust exploitation.</p><p><strong>Recommendation:</strong> <br>Strictly enforce Zero Trust architecture, combined with mandatory multi-party cross-verification mechanisms.</p><p><strong>Pain Point 2: Security environments are becoming highly dynamic — static defense is no longer effective</strong> <br>AI enables “instant mining” of historical vulnerabilities, rendering traditional one-time audits insufficient.</p><p><strong>Recommendation:</strong> <br>Projects should establish continuous, 24/7 auditing systems, integrating AI security platforms such as GoPlus DeepScan into CI/CD pipelines to achieve real-time monitoring and dynamic risk detection.</p><p><strong>Pain Point 3: Boundaries of AI supply chain and model access control are increasingly blurred</strong> <br>Incidents such as the Vercel supply chain attack, Anthropic Mythos unauthorized access, and agent-driven production database deletion demonstrate that prompt injection, credential leakage, and third-party toolchain poisoning are rapidly moving from the application layer into infrastructure layers.</p><p><strong>Recommendation:</strong> Adopt security tools such as AgentGuard to continuously scan agent toolchains and enforce runtime monitoring. In production environments, strictly isolate unknown plugins and dependencies.</p><p>In the era of deep integration between Web3 and AI, any blind confidence based on past security records can lead to catastrophic consequences.</p><p>Only by maintaining constant vigilance and continuously upgrading security awareness and tooling (GoPlus DeepScan + AgentGuard) can builders effectively defend against emerging and unknown threats.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1760df4a318e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GoPlus Launches AI Auditing Benchmark]]></title>
            <link>https://goplussecurity.medium.com/goplus-launches-ai-auditing-benchmark-94bf4f16fdc7?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/94bf4f16fdc7</guid>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Wed, 22 Apr 2026 10:17:48 GMT</pubDate>
            <atom:updated>2026-04-22T10:17:48.797Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>— Redefining the Boundaries of AI Smart Contract Auditing with Real-World Attack Data</strong></p><h4><strong>What is AI Auditing Benchmark？</strong></h4><p>The GoPlus AI Auditing Benchmark is a dataset designed for AI-driven smart contract auditing. Grounded in real-world attack incidents, it curates and extracts vulnerable contract source code from exploited projects, enabling systematic evaluation of AI’s real-world performance in vulnerability detection, attack path analysis, contextual reasoning, and auditing stability.</p><ul><li><strong>Core Focus: <br></strong>Whether AI auditing systems can identify vulnerabilities that lead to real financial loss in complex, real-world, context-rich code environments, and whether they can provide continuous auditing capability.</li><li><strong>Current Progress: <br></strong>Includes 22 real and representative smart contract attack incidents from May 2025 to April 2026, with ongoing updates.</li><li><strong>benchmark_complete: <br></strong>Preserves the full source tree of exploited contracts (business logic, dependencies, interfaces), suitable for evaluating performance across complex call flows, cross-module state, and full-context analysis.</li><li><strong>benchmark_simplified: <br></strong>Keeps only vulnerability-related functions and minimal dependencies, removing irrelevant logic for efficient detection, fast regression testing, and lower token usage and compute costs.</li></ul><h4>Get Involved</h4><p>If you are building in AI auditing, we invite you to explore, use, and contribute to the benchmark:<a href="https://github.com/GoPlusSecurity/ai-auditing-benchmark">https://github.com/GoPlusSecurity/ai-auditing-benchmark</a></p><h4>Vision</h4><p>Our goal is to build this into a <strong>representative public benchmark dataset</strong> for AI smart contract auditing — one that reflects real-world risk, real attack patterns, and real evaluation standards.</p><p>#GoPlus #Security #AI #SmartContract #DeepScan</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=94bf4f16fdc7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[️ From 512K Lines of Leaked Code: How Claude Code Achieves “Security First”]]></title>
            <link>https://goplussecurity.medium.com/%EF%B8%8F-from-512k-lines-of-leaked-code-how-claude-code-achieves-security-first-fb17ac52f4f5?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb17ac52f4f5</guid>
            <category><![CDATA[claude-code]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Wed, 01 Apr 2026 10:31:29 GMT</pubDate>
            <atom:updated>2026-04-01T10:31:29.939Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_UOGLloMjWZZhJjG6uKJtQ.png" /></figure><h3>I. Design Philosophy &amp; Security Architecture Overview</h3><p>Claude Code places <strong>“Security First” at the core of its design philosophy</strong>. This has been a defining characteristic since its inception, and a key factor that clearly differentiates it from OpenAI. It is also one of the main drivers behind Claude Code’s rapid growth in the enterprise market in recent years.</p><p>As an AI Agent capable of executing local code, reading and writing to the file system, initiating network requests, and even running high-risk terminal commands, it faces a threat surface far beyond traditional web applications or simple conversational LLMs.</p><p>These risks include code execution escape, path traversal, credential exfiltration, and prompt injection.</p><p>To address these challenges, Claude Code does not rely solely on regex-based allowlists or denylists. Instead, it builds a <strong>multi-layered, defense-in-depth architecture</strong>, spanning from the <strong>system kernel layer (OS Namespace)</strong> to the <strong>application layer (AST analysis and AI classifiers)</strong>, where each layer reinforces and backs up the others.</p><p>This system not only ensures strong protection of local assets, but also maintains a high level of developer experience and flexibility.</p><h3>II. Technical Analysis of the Defense-in-Depth System</h3><h4>1. Permission Layering &amp; Learning-based Fail-Open (Permission System)</h4><p><strong>Function Description:</strong><br> Claude Code’s permission system is not a binary “allow / deny” model. Instead, it is designed with three states: <strong>allow</strong>, <strong>deny</strong>, and <strong>ask</strong> (pause and request user confirmation).</p><p>With Auto Mode enabled, the system uses an AI classifier to dynamically evaluate and approve <strong>ask-level requests</strong>.</p><p>When the classifier becomes unavailable due to model degradation or network interruption, the system adopts a <strong>“learning-based fallback strategy” combined with a Fail-Open design</strong>, smoothly degrading into standard human-in-the-loop authorization via pop-up confirmation, rather than causing the entire toolchain to break.</p><p><strong>Code Implementation:</strong></p><ul><li>types/permissions.ts</li><li>utils/permissions/permissions.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// types/permissions.ts<br>export type PermissionBehavior = &#39;allow&#39; | &#39;deny&#39; | &#39;ask&#39;<br><br>// utils/permissions/permissions.ts<br>if (classifierResult.unavailable) {<br>  // Fail open: gracefully fall back to the standard permission flow, requiring manual user intervention<br>  logForDebugging(<br>    &#39;Auto mode classifier unavailable, falling back to normal permission handling (fail open)&#39;,<br>    { level: &#39;warn&#39; }<br>  )<br>  returnresult<br>}</pre><h4>2. Enterprise Hook Interception Chain &amp; Deny Rules</h4><p><strong>Function Description:</strong><br> In enterprise collaboration and internal security compliance scenarios, Claude Code provides a <strong>Hook interception chain</strong> (e.g. PreToolUse). It supports dynamically pulling enterprise <strong>policySettings</strong> from the backend and can enforce strict validation of enterprise-level <strong>Deny rules</strong>.</p><p>For example, enterprises can configure allowManagedHooksOnly to force-disable unsafe developer-defined Hooks, and forward commands to internal security auditing services for interception before they are executed by the BashTool.</p><p><strong>Code Implementation:</strong></p><ul><li>utils/hooks/hooksConfigSnapshot.ts</li><li>services/tools/toolHooks.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// utils/hooks/hooksConfigSnapshot.ts<br>functiongetHooksFromAllowedSources(): HooksSettings {<br>  const policySettings = settingsModule.getSettingsForSource(&#39;policySettings&#39;)<br><br>  if (policySettings?.disableAllHooks === true) return {}<br><br>  // Enforced control: only allow managed Hooks to run<br>  if (policySettings?.allowManagedHooksOnly === true) {<br>    return policySettings.hooks ?? {}<br>  }<br>  // ...<br>}</pre><h4>3. AI Classifiers for Prompt Injection Defense</h4><p><strong>Function Description:</strong><br> As Prompt Injection and obfuscated privilege-escalation attacks become increasingly sophisticated, traditional static analysis is becoming easier to bypass. Claude Code introduces a <strong>two-stage XML Classifier / YOLO Classifier</strong>.</p><p>When handling automated decisions, the AI evaluates the intent behind tool actions issued by the large model from a <strong>security perspective</strong> — such as whether they involve weakening system protections or attempting to disable Defender — effectively using <strong>“magic to defeat magic.”</strong></p><p><strong>Code Implementation:</strong></p><ul><li>utils/permissions/yoloClassifier.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// utils/permissions/yoloClassifier.ts<br>// Targeted deny guidance injection for privilege escalation scenarios<br>constpowershellDenyGuidance = getPlatform() === &#39;windows&#39; ? [<br>  &#39;PowerShell Elevation: `Start-Process -Verb RunAs`, `-ExecutionPolicy Bypass`, and disabling AMSI/Defender (`Set-MpPreference -DisableRealtimeMonitoring`) are classified under &quot;Security Weaken&quot;.&#39;<br>] : []</pre><h4>4. BashTool Sandbox: AST Analysis &amp; Iterative Unpacking</h4><p><strong>Function Description:</strong><br>Bash command execution is the highest-risk entry point. Claude Code builds an <strong>extremely strict filtering layer at the application level</strong>:</p><ol><li><strong>25 AST-based static checks</strong>: Using Tree-sitter to parse shell syntax, precisely defending against attacks such as JQ injection, carriage return (CR) injection, and comment/quote desynchronization.</li><li><strong>Fixed-Point Iterative Unpacking</strong>: Attackers often bypass prefix matching with patterns like FOO=bar nohup rm -rf /. The system uses a while loop to repeatedly strip environment variables and wrappers until the <strong>underlying real command is fully exposed</strong>, and only then performs permission checks.</li></ol><p><strong>Code Implementation:</strong></p><ul><li>tools/BashTool/bashSecurity.ts</li><li>tools/BashTool/bashPermissions.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// tools/BashTool/bashSecurity.ts (core validator chain)<br>constvalidators = [<br>  validateJqCommand,<br>  validateDangerousVariables,<br>  validateCommentQuoteDesync, // intercept comment/quote desynchronization vulnerabilities<br>  validateCarriageReturn, // intercept carriage return (CR) obfuscation<br>  // ... total of 25 checks<br>  validateMalformedTokenInjection,<br>]<br><br>// tools/BashTool/bashPermissions.ts (unpacking logic)<br>while (startIdx &lt; commandsToTry.length) {<br>  // iteratively strip layers until convergence<br>  const envStripped = stripAllLeadingEnvVars(cmd)<br>  const wrapperStripped = stripSafeWrappers(cmd)<br>  // ...<br>}</pre><h4>5. OS-Level Sandbox Isolation (OS-Level Sandbox)</h4><p><strong>Function Description:</strong><br>To prevent application-layer checks from being bypassed by 0-day vulnerabilities, Claude Code integrates a <strong>kernel-level isolation solution</strong> (@anthropic-ai/sandbox-runtime).</p><p>On Linux/WSL2, it leverages bwrap to establish <strong>Mount/PID Namespaces</strong>; on macOS, it uses sandbox-exec.</p><p>All subprocesses run within <strong>read-only or minimally privileged environments</strong>, forming the <strong>final physical line of defense</strong> against catastrophic damage caused by arbitrary code execution (RCE).</p><p><strong>Code Implementation:</strong></p><ul><li>utils/sandbox/sandbox-adapter.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// utils/sandbox/sandbox-adapter.ts<br>export const SandboxManager: ISandboxManager = {<br>  // intercept file read/write, restrict network access, and apply OS-level namespace isolation<br>  getFsReadConfig: BaseSandboxManager.getFsReadConfig,<br>  getNetworkRestrictionConfig: BaseSandboxManager.getNetworkRestrictionConfig,<br>  isSupportedPlatform: memoize((): boolean =&gt; BaseSandboxManager.isSupportedPlatform()),<br>  wrapWithSandbox,<br>}</pre><h4>6. Deep File System Defense (Path Traversal &amp; Unicode Normalization)</h4><p><strong>Function Description:</strong><br> To defend against path traversal attacks that could expose system files such as /etc/passwd, Claude Code goes beyond simply blocking traditional ../ patterns, and introduces protections targeting deeper attack vectors:</p><ul><li><strong>Null Byte (</strong><strong>\0) interception</strong>: Prevents C-level string truncation attacks.</li><li><strong>Unicode normalization defense (NFKC)</strong>: Blocks full-width characters (e.g. ．．／) that could be interpreted as path traversal at the underlying file system level (ASCII smuggling attacks).</li><li><strong>Symlink resolution interception</strong>: Uses realpath to prevent escape via symbolic links.</li></ul><p><strong>Code Implementation:</strong></p><ul><li>memdir/teamMemPaths.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// memdir/teamMemPaths.ts<br>functionsanitizePathKey(key: string): string {<br>  if (key.includes(&#39;\0&#39;)) throw new PathTraversalError(...)<br><br>  // NFKC normalization defense against full-width encoding exploits (e.g. U+FF0E, U+FF0F)<br>  constnormalized = key.normalize(&#39;NFKC&#39;)<br>  if (normalized !== key &amp;&amp; (normalized.includes(&#39;..&#39;) || normalized.includes(&#39;/&#39;))) {<br>    thrownewPathTraversalError(`Unicode-normalized traversal`)<br>  }<br>}</pre><h4>7. Network Boundary Control &amp; SSRF Protection</h4><p><strong>Function Description:</strong><br> SSRF attacks, where an Agent is induced to probe cloud metadata endpoints (such as 169.254.169.254), pose a critical security risk.</p><p>Claude Code enforces strict controls at the network layer:</p><ul><li>If a sandbox proxy is enabled, requests are intercepted at the underlying proxy layer.</li><li>If no proxy is enabled, a ssrfGuardedLookup hook is attached before the axios request is sent.</li></ul><p>The system resolves the target domain to its real IP address, and immediately blocks the connection if it detects:<br> → private internal network addresses<br> → loopback addresses</p><p><strong>Code Implementation:</strong></p><ul><li>utils/hooks/execHttpHook.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// utils/hooks/execHttpHook.ts<br>const response = await axios.post&lt;string&gt;(hook.url, jsonInput, {<br>  // SSRF guard: validate the resolved IP address and block private / link-local IP ranges<br>  lookup: sandboxProxy || envProxyActive ? undefined : ssrfGuardedLookup,<br>})</pre><h4>8. Windows / PowerShell-Specific Defense &amp; SMB Blocking</h4><p><strong>Function Description:</strong><br>Given the unique architecture of Windows, the system implements a dedicated <strong>AST analysis engine for PowerShellTool</strong>, specifically designed to defend against:</p><ul><li>.NET COM object abuse</li><li>malware downloads via Invoke-WebRequest</li><li>privilege escalation using Start-Process -Verb RunAs</li></ul><p>Most critically, it <strong>directly blocks all UNC path access (</strong><strong>\\Server\Share)</strong>.</p><p>This effectively prevents malicious prompts from inducing PowerShell to access remote SMB shares, thereby avoiding leakage of the system’s <strong>NTLM / Kerberos credentials</strong>.</p><p><strong>Code Implementation:</strong></p><ul><li>tools/PowerShellTool/pathValidation.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// tools/PowerShellTool/pathValidation.ts<br>// block UNC paths to prevent NTLM credential leakage via sniffing<br>if (normalizedPath.startsWith(&#39;//&#39;) || /DavWWWRoot/i.test(normalizedPath)) {<br>  return {<br>    allowed: false,<br>    decisionReason: {<br>      type: &#39;other&#39;,<br>      reason: &#39;UNC paths are blocked because they can trigger network requests and credential leakage&#39;<br>    }<br>  }<br>}</pre><h4>9. Local Secret Scanning &amp; MCP Key Isolation</h4><p><strong>Function Description:</strong><br> For credential management, <strong>preventing leakage and ensuring secure storage are equally critical</strong>:</p><ul><li><strong>Write &amp; outbound interception</strong>:<br> Integrates a lightweight version of <em>gitleaks</em> with high-confidence rules. Before content is synced to Team Memory or written to disk, it is statically scanned. If sensitive data such as AWS tokens or GitHub PATs are detected, they are either masked locally as [REDACTED] or the write operation is directly blocked.</li><li><strong>MCP plugin key isolation</strong>:<br> When users configure MCP plugins, the system leverages the sensitive attribute in the schema to separate data:<br> → non-sensitive configs (e.g. URLs) are stored in plaintext in settings.json<br> → high-risk tokens are <strong>forcefully stored in OS-level secure storage</strong> (macOS Keychain or Linux credential files with 0600 permissions), completely eliminating the risk of unauthorized access</li></ul><p><strong>Code Implementation:</strong></p><ul><li>services/teamMemorySync/secretScanner.ts</li><li>utils/plugins/mcpbHandler.ts</li></ul><p><strong>Key Code Analysis:</strong></p><pre>// utils/plugins/mcpbHandler.ts<br>for (const [key, value] of Object.entries(config)) {<br>  // rely on the schema to force isolation of sensitive data into the Keychain<br>  if (schema[key]?.sensitive === true) {<br>    sensitive[key] = String(value)<br>  } else {<br>    nonSensitive[key] = value<br>  }<br>}</pre><h3>III. Summary</h3><p>Claude Code’s security capabilities go beyond the vast majority of current Agent frameworks that rely on single-point filtering.</p><p>At the core of its architecture is a <strong>“zero trust” design philosophy</strong> — <em>never trusting any single layer of defense</em>:</p><ol><li>Models can make mistakes → therefore AST analysis and classifiers are in place</li><li>AST analysis can be bypassed → therefore an underlying OS-level sandbox is enforced</li><li>Beyond the sandbox → there are dedicated defenses targeting specific attack vectors (such as SSRF, Unicode smuggling, and NTLM leakage)</li><li>Even with user-configured setups → sensitive data is still isolated via OS-level key management (e.g. Keychain), preventing unauthorized access</li></ol><p>This multi-dimensional security design — combining<br><strong>static rules + AI-driven detection + OS-level isolation + protocol-level blocking</strong> —</p><p>demonstrates Claude Code’s <strong>“Security First” philosophy</strong>, balancing enterprise-grade security, adversarial resilience, and control over the boundaries of developer tooling.</p><p>👉 This is a security architecture worth studying for every AI Agent system.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fb17ac52f4f5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ClawHub Top 100 Skills Security Scan — In-Depth Analysis Report]]></title>
            <link>https://goplussecurity.medium.com/clawhub-top-100-skills-security-scan-in-depth-analysis-report-ac47f7ef6f3d?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/ac47f7ef6f3d</guid>
            <category><![CDATA[openclaw]]></category>
            <category><![CDATA[security]]></category>
            <category><![CDATA[skills]]></category>
            <category><![CDATA[bnb-chain]]></category>
            <category><![CDATA[agentguard]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Thu, 12 Mar 2026 10:36:39 GMT</pubDate>
            <atom:updated>2026-03-12T10:36:39.049Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dt3rXPUcZNAXSVZZ4zp3aQ.jpeg" /></figure><p><strong>Report Date</strong>: 2026–03–11</p><p><strong>Sample Scope</strong>: ClawHub Top 100 Skills by downloads (as of March 10, 18:30 Beijing Time)</p><p><strong>Detection Engine</strong>: AgentGuard Skills Security Scanning Engine</p><p><strong>Purpose of the Scan</strong>: To evaluate the baseline security posture of the most popular AI Agent Skills and identify potential risks such as privilege abuse, sensitive operations, and malicious behavior patterns.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*omo3r6wBAD1FNufyHr3CXQ.png" /></figure><h3>📊 1. Executive Summary</h3><p>This security scan conducted a full analysis of the <strong>100 most frequently downloaded Skills in the ClawHub ecosystem</strong>. The overall results are as follows:</p><ul><li><strong>Total Samples Scanned</strong>: 100</li><li><strong>Successful Scan Rate</strong>: 100% (no parsing failures or missing files)</li><li><strong>Blocked Skills</strong>: 21 (21%)</li><li><strong>Warning Skills</strong>: 17 (17%)</li><li><strong>Passed Skills</strong>: 62 (62%)</li></ul><p><strong>Key Finding</strong>: Among the Top 100 Skills, <strong>21% contain explicit high-risk operations</strong> (such as direct network tunneling, sensitive API calls, or automated messaging). For these Skills, it is recommended to enforce a <strong>Human-in-the-Loop (HITL)</strong> confirmation mechanism before execution to ensure manual review of high-risk actions</p><p>Additionally, <strong>17% of Skills present certain risk signals</strong>, and should be executed with caution. For users with stricter security requirements, enabling manual confirmation for these Skills is also recommended.</p><h3>📈 2. Risk Level Distribution</h3><p>Based on the <strong>AgentGuard rule set</strong>, the scan results are categorized into <strong>four risk levels</strong>, with the following distribution:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EI6BmCdp-5uRRBhS5amG8Q.jpeg" /></figure><h3>3. High-Risk (Critical &amp; High) Skills Analysis</h3><p>In this security scan, <strong>21 Skills were directly classified as <em>Blocked</em></strong> after triggering <strong>Critical or High-risk rules</strong>. These Skills mainly fall into the following high-risk operational scenarios:</p><p><strong>3.1. Headless Browsers &amp; Automation (Browser Automation)</strong></p><p>These Skills typically call <strong>Puppeteer/Playwright</strong> or packaged CLI tools, allowing the Agent to freely access the internet.</p><ul><li><strong>Affected Skills</strong>: agent-browser (Critical), agent-browser-clawdbot (Critical)</li><li><strong>Detection Reason</strong>: Detection of headless browser process control, execution of arbitrary JS scripts, or complex DOM interactions.</li><li><strong>Risk Impact</strong>: May be used for <strong>SSRF (Server-Side Request Forgery)</strong>, internal network probing, bypassing CAPTCHA protections for malicious scraping, or triggering payloads from phishing websites.</li></ul><p><strong>3.2. Communications &amp; Messaging (Communications &amp; Messaging)</strong></p><p>Skills that directly control users’ email or messaging tools to send messages.</p><ul><li><strong>Affected Skills</strong>: agentmail (Critical), whatsapp-business (Critical), imap-smtp-email (Critical), mailchimp (High)</li><li><strong>Detection Reason</strong>: Detection of SMTP/IMAP protocol keywords, bulk messaging API endpoints, and high-privilege communication tokens.</li><li><strong>Risk Impact</strong>: If an Agent is compromised through a <strong>Prompt Injection attack</strong>, attackers could use these Skills to send spam or conduct social engineering scams, damaging user reputation and potentially violating privacy compliance requirements.</li></ul><p><strong>3.3. High-Privilege CRM &amp; Cloud Resource APIs (Enterprise API Gateways)</strong></p><p>Skills that perform read/write operations on enterprise SaaS platforms through proxy gateways.</p><ul><li><strong>Affected Skills</strong>: google-workspace-admin (Critical), google-slides (Critical), feishu-evolver-wrapper (Critical), pipedrive-api (High), youtube-api-skill (High), trello-api (High), google-meet (High)</li><li><strong>Detection Reason</strong>: The rule engine detected REST API request structures capable of modifying sensitive data, as well as the ability to alter the state of external systems.</li><li><strong>Risk Impact</strong>: Agents may directly modify or even delete core enterprise data such as sales, financial, document, or media assets (including granting or revoking administrator privileges). Without strict <strong>Human-in-the-Loop verification</strong>, operational mistakes could lead to severe losses.</li></ul><p><strong>3.4. Deep Search Engines &amp; Scraping Aggregation (Search &amp; Scraping)</strong></p><p>Tools capable of deep content crawling and multi-engine aggregation.</p><ul><li><strong>Affected Skills</strong>: brave-search (High), duckduckgo-search (High), multi-search-engine (High), tavily (High)</li><li><strong>Detection Reason</strong>: Detection of high-frequency external network request wrappers, HTML parsing, and scraping libraries.</li><li><strong>Risk Impact</strong>: In addition to potential IP bans from target websites, unsafe web content extraction (such as directly reading unfiltered HTML or executing webpage content) may introduce <strong>Indirect Prompt Injection</strong> risks.</li></ul><p><strong>3.5. Core Logic Modification &amp; Privilege Escalation (Self-Modification &amp; Privilege Escalation)</strong></p><p>Skills involving Agent behavior mutation, modification of hidden configurations, or access to system-level credentials.</p><ul><li><strong>Affected Skills</strong>: free-ride (Critical), moltbook-interact (Critical), trello (Critical), evolver (High)</li><li><strong>Detection Reason</strong>: Detection of write access requests to system files (such as hidden .json configuration files), the ability to read local credentials, or direct concatenation of sensitive tokens in Bash commands.</li><li><strong>Risk Impact</strong>: These Skills may modify the Agent’s core behavioral rules without authorization, bypass existing system restrictions, or leak critical API keys due to command injection vulnerabilities.</li></ul><h3>⚠️ 4. Medium-Risk (Medium / Warning) Skills Worth Attention</h3><p>A total of <strong>17 Skills were classified as Medium risk</strong>. Most of these are integration interfaces for third-party applications. While the scanning engine did not directly block them, the following considerations apply:</p><ul><li><strong>Calendar &amp; scheduling tools</strong> (caldav-calendar, calendly-api)</li><li><strong>Productivity &amp; collaboration tools</strong> (notion, x-twitter, xero, typeform)</li><li><strong>Microsoft ecosystem integrations</strong> (microsoft-excel, outlook-graph, outlook-api)</li><li><strong>Utility toolkits</strong> (mcporter, asana-api, clickup-api)</li></ul><p><strong>Analysis</strong>: The key characteristic of these Skills is that <strong>their functionality is neutral, but they carry high-value credentials</strong>. The security scan identified that they require access tokens from users, but no explicit malicious logic was detected at the code level, so they were labeled with a Warning. However, if an Agent is <strong>tricked into using these Skills</strong> (for example, being instructed to “delete all my meetings tomorrow” or “publicly share this Notion document”), significant damage could still occur.</p><h3>🛡️ 5. Summary &amp; Security Recommendations</h3><p>This security scan shows that <strong>around 20% of the Top 100 most-downloaded Skills on ClawHub contain explicit high-risk operations</strong>, demonstrating the necessity of performing security scanning and review for Skills.</p><p><strong>Recommendations for users and ecosystem developers:</strong></p><p><strong>5.1. Improve the usability of high-risk Skills (Blocked)</strong>: Avoid blanket bans. For high-value Skills such as <strong>agent-browser</strong> and <strong>agentmail</strong>, it is recommended to enforce a <strong>Human-in-the-Loop (HITL)</strong> confirmation mechanism before execution. The specific content to be sent or actions to be performed should be displayed, and execution should proceed only after explicit user authorization.</p><p><strong>5.2. Strengthen protection against “Indirect Prompt Injection”</strong>: For all search-related Skills marked as <strong>High</strong>, the content returned to the Agent must undergo strict sanitization (such as stripping HTML tags and scripts) to prevent malicious content injection from external webpages.</p><p><strong>5.3. Conduct regular security health checks</strong>: Based on the security blind spots identified in this analysis, actions such as <strong>modifying specific configuration file paths (e.g., AGENTS.md)</strong> and <strong>invoking system desktop control libraries</strong> should be added to the high-risk behavior list. Alternatively, developers can use <strong>AgentGuard security inspection and Skills scanning capabilities</strong> to regularly audit the security posture of Agents and their Skills.</p><h3>👉<strong>Appendix: Detailed Results of the ClawHub Top 100 Skills Security Scan</strong></h3><p>Click the link to view:</p><p><a href="https://inky-punch-9d2.notion.site/Appendix-Detailed-Results-of-the-ClawHub-Top-100-Skills-Security-Sca-3215da0dd7ad80719937c66b7c1225b3?source=copy_link">https://inky-punch-9d2.notion.site/Appendix-Detailed-Results-of-the-ClawHub-Top-100-Skills-Security-Sca-3215da0dd7ad80719937c66b7c1225b3?source=copy_link</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ac47f7ef6f3d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Who Is Stealing Your Crypto Assets? — 2025 Web3 User Security and Risk Trends Report]]></title>
            <link>https://goplussecurity.medium.com/who-is-stealing-your-crypto-assets-2025-web3-user-security-and-risk-trends-report-1deb46dac16c?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/1deb46dac16c</guid>
            <category><![CDATA[goplus]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Tue, 30 Dec 2025 03:08:20 GMT</pubDate>
            <atom:updated>2025-12-30T03:08:20.110Z</atom:updated>
            <content:encoded><![CDATA[<h3>Who Is Stealing Your Crypto Assets? — 2025 Web3 User Security and Risk Trends Report</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yQX39wPtWtZUSYAJ0jkdOQ.png" /></figure><h3>Chapter 1: A Grim Security Landscape in 2025</h3><p>In 2025, according to data from the <strong>GoPlus RektDatabase</strong>, the Web3 space experienced over <strong>1,200</strong> major security incidents involving users and project teams, resulting in total losses exceeding <strong>USD 3.5 billion</strong>.</p><p>The overall security landscape shows two clear trends: an <strong>increase in the number of ultra-large-scale incidents</strong> and a <strong>significant reduction in the cost of conducting small-scale fraud against users</strong>. This indicates that attackers are simultaneously adopting strategies of <strong>“precision targeting”</strong> and <strong>“wide-net exploitation.”</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LuLYMB0SwMMr5IqJtqIfYw.png" /></figure><h3>1.1 Top Loss Incidents of 2025 (Precision Targeting Attacks)</h3><ul><li><strong>February 21</strong> — <strong>Bybit</strong> was attacked, resulting in losses of approximately <strong>USD 1.5 billion</strong>. The cause was the compromise of a Safe multisig server, where transaction instructions were tampered with. <strong>Follow-up actions:</strong> Bybit cooperated with law enforcement agencies including the FBI, freezing over <strong>USD 40 million</strong> in stolen assets. The platform replenished reserves to ensure that user funds were not affected.</li><li><strong>May 22</strong> — <strong>Cetus Protocol</strong> was attacked, suffering losses of approximately <strong>USD 223 million</strong> on <strong>Sui</strong>, caused by a vulnerability in a Move smart contract. The price of <strong>$CETUS</strong> dropped by more than <strong>16%</strong>. <strong>Follow-up actions:</strong> The Sui network quickly coordinated to freeze <strong>USD 162 million</strong> of the stolen assets. A community vote approved a <strong>100% compensation plan</strong>, with additional <strong>USD 30 million</strong> in emergency credit support from the Sui Foundation.</li><li><strong>November 3</strong> — <strong>Balancer</strong> was attacked, with multi-chain losses of approximately <strong>USD 128 million</strong>, due to a vulnerability in the Balancer V2 smart contract. <strong>Follow-up actions:</strong> Berachain coordinated validators to execute an emergency hard fork and freeze the attacker’s addresses. White-hat operators returned <strong>USD 12.8 million</strong>, and StakeWise recovered approximately <strong>USD 20 million</strong> using contract management mechanisms.</li><li><strong>June 18</strong> — <strong>Iranian exchange Nobitex</strong> was attacked, resulting in multi-chain losses of approximately <strong>USD 90 million</strong>. <strong>Follow-up actions:</strong> This incident had clear geopolitical overtones. The attacker transferred funds to unusable, humiliating addresses, aiming to cause operational paralysis rather than direct financial gain.</li><li><strong>April 1</strong> — <strong>UPCX</strong> was attacked, with losses of approximately <strong>USD 70 million</strong>, caused by leaked administrator private keys that allowed unauthorized upgrades to the ProxyAdmin contract. <strong>Follow-up actions:</strong> Remaining tokens were transferred, and law enforcement agencies were engaged to track the stolen assets.</li><li><strong>January 23</strong> — <strong>Phemex</strong> was attacked, suffering multi-chain losses of approximately <strong>USD 69.1 million</strong>, due to a compromised hot wallet. <strong>Follow-up actions:</strong> The platform released Proof of Reserves (PoR) to confirm cold wallet security and compensated affected users in full through phased reimbursements using its own funds.</li><li><strong>August 14</strong> — <strong>Turkish exchange BtcTurk</strong> was attacked, with losses of approximately <strong>USD 54 million</strong>, caused by a hot wallet breach. This marked the second hot wallet theft for the exchange within <strong>14 months</strong>. <strong>Follow-up actions:</strong> A <strong>USD 2.4 million</strong> bounty program was launched to solicit leads.</li><li><strong>December 20</strong> — A <strong>high-net-worth investor address (0xcB8078)</strong> suffered an <strong>address poisoning phishing attack</strong>, resulting in a loss of <strong>USD 49.99 million in USDT</strong>. <strong>Follow-up actions:</strong> The victim issued an on-chain message offering a <strong>USD 1 million white-hat bounty</strong> in exchange for the return of <strong>98%</strong> of the funds. However, the attacker has since swapped the assets into ETH and laundered them via <strong>Tornado Cash</strong>.</li><li><strong>February 24</strong> — <strong>Infini</strong> was attacked, with losses of approximately <strong>USD 49.5 million</strong>, due to flaws in administrator permission management. <strong>Follow-up actions:</strong> The founder issued a public apology and established a dedicated compensation fund, committing to <strong>100% reimbursement</strong> for all affected users by <strong>2026</strong>.</li><li><strong>July 20</strong> — <strong>Indian exchange CoinDCX</strong> was attacked, with losses of approximately <strong>USD 44.2 million</strong>, caused by a hot wallet compromise. <strong>Follow-up actions:</strong> A user compensation plan was announced and efforts to trace the stolen assets were initiated.</li><li><strong>July 9</strong> — <strong>GMX</strong> was attacked, losing approximately <strong>USD 40.5 million</strong> due to a smart contract vulnerability. <strong>Follow-up actions:</strong> Within <strong>48 hours</strong>, the attacker accepted a <strong>10% white-hat bounty</strong> (approximately <strong>USD 5 million</strong>) proposed by the project and returned around <strong>USD 40 million</strong> of the stolen funds.</li><li><strong>December 2</strong> — <strong>South Korean exchange Upbit</strong> was attacked, suffering losses of approximately <strong>USD 30 million</strong> on <strong>Solana</strong>, attributed to infiltration by a <strong>North Korean hacking group</strong>. <strong>Follow-up actions:</strong> Upbit cooperated with global exchanges to implement real-time blacklist tagging and tracking/interception of funds linked to the attackers.</li></ul><p><strong>CeFi as the Primary Hacker Cash-Out Machine:</strong></p><p>In 2025, there were <strong>12 attack incidents</strong> with single-event losses exceeding <strong>USD 30 million</strong>, <strong>7 of which involved CeFi platforms</strong>. The most common causes were <strong>compromised administrator private keys</strong> and <strong>hot wallet private key theft</strong>, exposing significant systemic risks within centralized infrastructures.</p><p><strong>DeFi Vulnerabilities Are Becoming Deeper and More Sophisticated:</strong></p><p>In 2025, the DeFi sector demonstrated stronger resilience compared to previous years. Losses caused by smart contract vulnerabilities grew at a rate <strong>far lower than the growth of TVL</strong>. However, it is worth noting that long-running and previously secure protocols such as <strong>Balancer</strong> and <strong>Yearn</strong> continue to see <strong>new vulnerabilities discovered</strong>, and <strong>Move-based smart contract exploits</strong> have increased significantly. This reflects the rapidly improving ability of attackers to leverage <strong>AI technologies</strong> to discover and exploit smart contract weaknesses.</p><p><strong>The Peak “Performance” of State-Sponsored Hackers:</strong></p><p>In 2025, <strong>North Korean hacking groups</strong> (such as the <strong>Lazarus Group</strong>) continued their rapid expansion in stolen funds, stealing at least <strong>USD 2.02 billion</strong> in crypto assets during the year — an increase of <strong>51% compared to 2024</strong>, marking a historical record for North Korea’s cyber-enabled asset acquisition.</p><p>In 2025, North Korean hacking operations underwent a <strong>deep evolution</strong> in tactics:</p><ul><li><strong>Social Engineering:</strong> By impersonating recruiters from Web3 or AI companies, hackers maintained relationships with targets for months on platforms such as LinkedIn, ultimately诱cing victims during “technical interviews” to download test code embedded with backdoors.</li><li><strong>Precision Targeting:</strong> Efforts were concentrated on a small number of institutions or individuals holding massive capital reserves. In 2025, <strong>North Korean hackers accounted for over 76% of the total losses</strong> in all institution-level security incidents.</li><li><strong>Industrialized Money Laundering:</strong> Once stolen, funds rapidly entered a highly automated laundering pipeline, typically completing the full process — from on-chain mixing to fiat settlement via illicit platforms — <strong>within 45 days</strong>.</li></ul><h3>1.2 High-Frequency Attacks and Fraud Types (Wide-Net Exploitation)</h3><p>According to data from the <strong>GoPlus RektDatabase</strong>, among Web3 security incidents in 2025, the three most frequent attack and fraud types were <strong>private key theft</strong> (via malware and social engineering), <strong>phishing attacks</strong>, and <strong>rug tokens</strong> (fraudulent tokens).</p><h3>Private Key Theft</h3><p>In 2025, private key theft fully shifted from early-stage “broad, indiscriminate attacks” to <strong>highly targeted hunting</strong>. Attackers increasingly focused on <strong>individual accounts with multisig permissions</strong> or <strong>centralized service providers</strong>.</p><ul><li><strong>Number of incidents:</strong> Over <strong>300 cases</strong> throughout the year.</li><li><strong>Total losses:</strong> Exceeded <strong>USD 1.8 billion</strong> (accounting for <strong>more than 50%</strong> of total annual losses).</li><li><strong>Key characteristics:</strong></li></ul><ol><li><strong>Malware-based attacks:</strong> Attackers disguised malicious software as popular Web3 productivity tools, node maintenance scripts, or DeFi analytics plugins, seducing users to download them and implanting <strong>private key stealers</strong>.</li><li><strong>Social engineering:</strong> A typical example is the <strong>“fake Zoom” social engineering attack</strong>. Attackers impersonated recruiters or project teams, invited core developers to interviews, and then delivered malicious “coding test tasks” or “meeting plugins” containing backdoors to gain device control and steal private keys.</li><li><strong>Supply chain compromise:</strong> Targets shifted from <strong>wallet files</strong> to <strong>development environments</strong>. In 2025, incidents caused by compromised development setups or the installation of malicious third-party dependencies leading to private key leakage increased significantly.</li></ol><h3>Phishing Attacks</h3><p>Phishing attacks remained the <strong>most frequent</strong> threat type in 2025, and with the widespread adoption of AI technologies, the number of fraud cases grew rapidly.</p><ul><li><strong>Number of incidents:</strong> Over <strong>400 cases</strong> throughout the year (large-scale, recorded incidents; small retail phishing cases are difficult to quantify).</li><li><strong>Total losses:</strong> Exceeded <strong>USD 750 million</strong>.</li><li><strong>Trend analysis:</strong></li></ul><ol><li><strong>AI-generated phishing:</strong> Attackers used large language models to automatically generate multilingual phishing emails and posts that were highly tailored to the preferences and behaviors of target users.</li><li><strong>Cross-platform expansion:</strong> Phishing vectors expanded from <strong>X (Twitter)</strong> to <strong>Telegram</strong>, <strong>Discord</strong>, and even <strong>LinkedIn</strong>.</li><li><strong>Signature traps:</strong> In 2025, <strong>signature-based phishing</strong> remained the dominant method, with attackers seduction users to sign malicious approvals and subsequently draining their assets.</li><li><strong>Defense upgrades:</strong> Real-time detection by security extensions and wallet-side <strong>transaction preview</strong> features significantly reduced the success rate of low-level phishing, pushing attackers toward more advanced and interactive fraud techniques.</li></ol><h3>Rug Tokens (Malicious Liquidity Pulls / Exit Scams)</h3><p>In 2025, rug tokens showed a clear shift from <strong>frequent small-scale scams</strong> to <strong>single, large-scale exits</strong>, especially as the boom in the meme market gave rise to more sophisticated extraction techniques.</p><ul><li><strong>Number of incidents:</strong> Approximately <strong>300 cases</strong> throughout the year (with high concentration on <strong>Base</strong>, <strong>BSC</strong>, and <strong>Solana</strong>).</li><li><strong>Total losses:</strong> Exceeded <strong>USD 600 million</strong> (excluding certain disputed price collapses).</li><li><strong>Key characteristics:</strong></li></ul><ol><li><strong>Fragmented rug pulls:</strong> Developers no longer remove all liquidity at once. Instead, they use large numbers of related addresses (<strong>puppet wallets</strong>) to withdraw small amounts over several days, avoiding monitoring alerts.</li><li><strong>Hidden backdoors:</strong> Malicious teams hide backdoors via <strong>proxy contracts</strong>. After passing initial audits, they exploit upgrade mechanisms to modify contract logic, insert malicious functions, or grant super-admin privileges — executing a “technical exit.”</li><li><strong>Professionalization and syndication:</strong> An underground industry offering <strong>“Rug Pull as a Service”</strong> has emerged, providing end-to-end solutions including code development, fake marketing, and volume manipulation.</li><li><strong>Meme coins as the main battlefield:</strong> Over <strong>90%</strong> of rug tokens occurred in the meme coin sector, which lacks real business fundamentals.</li></ol><h3>Chapter 2: Typical Cases and Attack Technique Breakdown</h3><h3>2.1 Bybit — An Epic Multisig Frontend Tampering Attack</h3><p>This was the <strong>most devastating security incident in the global Web3 space in 2025</strong>, a textbook <strong>supply-chain attack</strong>. Instead of attacking the target directly, the attackers compromised the multisig infrastructure it relied on — <strong>Safe{Wallet}</strong>.</p><ul><li><strong>Technical Breakdown &amp; Attack Flow</strong></li></ul><ol><li><strong>Frontend compromise:</strong> The North Korean hacking group <strong>Lazarus Group</strong> compromised the device of a Safe{Wallet} developer who had system release permissions. They then gained access to Safe’s AWS account and injected malicious scripts into the official frontend code.</li><li><strong>Targeted fraud:</strong> The malicious script was activated <strong>only for Bybit’s specific multisig wallet addresses</strong>. When Bybit’s three multisig administrators used the compromised official Safe frontend for routine fund transfers, the interface appeared normal, but the underlying transaction data was secretly replaced with calls to a malicious contract pre-deployed by the attackers.</li><li><strong>Blind signing vulnerability:</strong> When administrators signed transactions using hardware wallets, the devices were unable to fully parse and display the complex contract call data, resulting in <strong>“blind signing”</strong> and failure to detect the abnormal transaction.</li><li><strong>Control takeover:</strong> Through a DelegateCall, the malicious transaction upgraded the wallet’s logic contract to one controlled by the attacker, granting full control over the multisig wallet and enabling the transfer of massive amounts of assets.</li></ol><ul><li><strong>Security Recommendations</strong></li></ul><ol><li><strong>For project teams:</strong> Operations involving large sums of assets should implement <strong>independent, multi-party verification mechanisms</strong>, avoiding reliance on a single frontend or service for all signers.</li><li><strong>For users:</strong> When signing critical transactions, always carefully verify the <strong>transaction hash and detailed call data</strong> on wallet devices, and remain highly cautious of any <strong>blind signing</strong> requests.</li></ol><ul><li><strong>Further Reading</strong></li></ul><ol><li><a href="https://www.nccgroup.com/research-blog/in-depth-technical-analysis-of-the-bybit-hack/">https://www.nccgroup.com/research-blog/in-depth-technical-analysis-of-the-bybit-hack/</a></li><li><a href="https://coinacademy.fr/wp-content/uploads/2025/02/Bybit-Incident-Investigation-Report.pdf">https://coinacademy.fr/wp-content/uploads/2025/02/Bybit-Incident-Investigation-Report.pdf</a></li><li><a href="https://announcements.bybit.com/article/incident-update-unauthorized-activity-involving-eth-cold-wallet-blt292c0454d26e9140/?referrer=grok.com">https://announcements.bybit.com/article/incident-update-unauthorized-activity-involving-eth-cold-wallet-blt292c0454d26e9140/?referrer=grok.com</a></li></ol><h3>2.2 UXLINK — AI Deepfake &amp; Social Engineering Attack</h3><p>This case shows that <strong>humans remain the weakest security link</strong>, even with multisig.</p><ul><li><strong>Analysis &amp; Process:</strong></li></ul><ol><li><strong>AI Deepfake Trust:</strong> Attack began before Sept 22; attackers impersonated trusted partners on Telegram, using video calls with real-time face and voice cloning to gain trust.</li><li><strong>Trojan Implant:</strong> During calls, personnel clicked malicious links, implanting malware, leaking multiple multisig keys.</li><li><strong>Contract Takeover:</strong> Attackers removed original admins and added themselves using stolen keys.</li><li><strong>Theft &amp; Minting:</strong> Existing wallet assets were stolen; attackers then maliciously minted UXLINK tokens and sold on market.</li></ol><ul><li><strong>Security Recommendations:</strong></li></ul><ol><li><strong>Permission Isolation:</strong> Separate <strong>fund wallets</strong> from <strong>permission wallets</strong>; avoid holding significant funds in permission wallets and physically isolate signing devices.</li><li><strong>Personnel Security:</strong> Continuous security awareness training; vigilance against social engineering.</li><li><strong>Governance Upgrades:</strong> Update and replace deployed contracts and admin addresses to mitigate risks.</li></ol><ul><li><strong>More Details:</strong></li><li><a href="https://x.com/UXLINKofficial/status/1998228868365529555">https://x.com/UXLINKofficial/status/1998228868365529555</a></li></ul><h3>2.3 Balancer — Computational Precision &amp; Smart Contract Vulnerability</h3><p>This incident demonstrates how <strong>tiny errors in complex business logic and mathematical calculations</strong> in DeFi protocols can trigger massive risks and losses.</p><ul><li><strong>Analysis &amp; Process:</strong></li></ul><ol><li><strong>Root Cause:</strong> The Balancer attack stemmed from a <strong>precision issue (rounding down)</strong>.</li><li>During swap calculations in the Balancer Vault, results were rounded down, affecting token prices inside the Vault.</li><li>batchSwap amplified this vulnerability: by carefully crafting batch swap parameters, attackers could significantly depress Vault token prices and profit from the resulting price difference.</li></ol><ul><li><strong>Security Recommendations:</strong></li></ul><ol><li><strong>Continuous Deep Business Logic Audits:</strong> Use more advanced and efficient security techniques such as AI and formal verification to improve vulnerability discovery, especially for deep business logic issues that cannot rely solely on manual audits.</li></ol><ul><li><strong>More Details:</strong></li></ul><ol><li><a href="https://x.com/Balancer/status/1990856260988670132">https://x.com/Balancer/status/1990856260988670132</a></li><li><a href="https://x.com/GoPlusSecurity/status/1985371222230884799">https://x.com/GoPlusSecurity/status/1985371222230884799</a></li><li><a href="https://x.com/GoPlusZH/status/1985360069031289085">https://x.com/GoPlusZH/status/1985360069031289085</a></li></ol><h3>Chapter 3: Evolution of Attack Methods and Emerging Tech Scams in 2025</h3><p>In 2025, the evolution of Web3 security threats showed <strong>significant technological convergence, intelligence, and automation</strong>. Attackers no longer rely on a single technique but deeply integrate new protocol features, supply chain vulnerabilities, AI, and automated tools to create attacks that are <strong>more destructive, stealthy, and cost-effective</strong>.</p><h3>3.1 EIP-7702 and New Phishing Attacks</h3><p>EIP-7702, a core proposal of Ethereum’s Pectra upgrade, aims to <strong>enable smart EOAs</strong>, enhancing account abstraction and user experience. However, this innovation was quickly weaponized by hackers, spawning phishing attacks with <strong>lower barriers to entry and higher stealth</strong>.</p><p><strong>Mechanism:</strong></p><p>The core mechanism of EIP-7702 allows EOAs to temporarily inject code via the SetCode operation to enable <strong>delegated authorization</strong>. Attackers exploit this by pre-controlling leaked private key addresses and authorizing them to malicious Delegator contracts, which automatically transfer any funds the user later deposits.</p><p>Unlike traditional phishing, this model <strong>does not rely on frontend UI fraud</strong>, and conventional malicious authorization or signature detection mechanisms fail, further amplifying risk. Users fall prey to the “private key equals assets” misconception, leading to losses. Combined with <strong>batch execution (</strong><strong>Batch)</strong>, the efficiency of phishing attacks is significantly increased.</p><p><strong>EIP-7702 New Phishing Typically Follows Four Stages:</strong></p><ol><li><strong>Preparation:</strong> Attackers obtain private keys from the dark web or historical leaks and authorize them to malicious Delegator contracts deployed across multiple chains (commonly ETH, BSC, Base).</li><li><strong>Inducement:</strong> Attackers use DMs or social media to expose users to these private keys, making them believe it’s a “free gain.”</li><li><strong>Execution:</strong> Once the user deposits Gas or tokens, the malicious Delegator contract automatically withdraws the funds, causing immediate loss.</li><li><strong>Laundering:</strong> Funds are automatically transferred via mixers or cross-chain bridges, making tracking difficult.</li></ol><p><strong>Further Cases:</strong></p><p>In Q2 2025, the Inferno Drain group exploited EIP-7702 phishing attacks on a massive scale, stealing millions in crypto assets, including individual cases exceeding $1.5 million. More details:</p><ul><li><a href="https://x.com/GoPlusZH/status/1930805446262284771">https://x.com/GoPlusZH/status/1930805446262284771</a></li><li><a href="https://x.com/GoPlusZH/status/1946113759938019701">https://x.com/GoPlusZH/status/1946113759938019701</a></li><li><a href="https://x.com/GoPlusZH/status/1942854235542245386">https://x.com/GoPlusZH/status/1942854235542245386</a></li><li><a href="https://x.com/GoPlusSecurity/status/1930614660996518198">https://x.com/GoPlusSecurity/status/1930614660996518198</a></li><li><a href="https://x.com/GoPlusSecurity/status/1946510327312797930">https://x.com/GoPlusSecurity/status/1946510327312797930</a></li><li><a href="https://x.com/GoPlusSecurity/status/1942916699579421117">https://x.com/GoPlusSecurity/status/1942916699579421117</a></li></ul><h3>2. Software Supply Chain Attacks: Self-Propagating Worms &amp; Cross-Ecosystem Damage</h3><p>In 2025, open-source supply chain attacks surged. Worm-like malware such as <strong>Shai-Hulud</strong> and its 2.0 variant caused a <strong>global impact</strong> on ecosystems like npm and Maven, affecting over <strong>25,000 GitHub repositories</strong>.</p><p><strong>Mechanism:</strong></p><p>Unlike single-package poisoning, Shai-Hulud uses a <strong>self-replicating worm mechanism</strong>: after compromising a maintainer’s account, it injects postinstall/preinstall scripts to steal credentials and propagate automatically. It leverages <strong>TruffleHog</strong> to scan the file system and steal AWS/GCP/Azure keys, npm tokens, and GitHub PATs. The worm then traverses the dependency graph, injects malicious payloads, publishes new versions, and spreads exponentially.</p><p><strong>Software supply chain attacks generally follow three stages:</strong></p><ol><li><strong>Initial Compromise:</strong> Attackers steal developer account credentials for popular open-source projects (e.g., debug, chalk) via phishing or other methods and inject malicious code into repositories.</li><li><strong>Worm Behavior:</strong> Malicious scripts (e.g., setup_bun.js) execute during installation (preinstall), automatically stealing npm tokens and GitHub PATs from the developer’s environment. Using these credentials, the worm scans all other repositories owned by the developer, injects malicious code, and republishes, achieving cross-project infection.</li><li><strong>Execution &amp; Theft:</strong> Once other developers use the infected repositories, attackers perform remote loading and background execution to steal funds or cause damage.</li><li><strong>Cross-Ecosystem Propagation:</strong> The infection spreads from npm to Maven/Java, with attack windows ranging from hours to days, affecting <strong>over 1 billion downloads</strong>.</li></ol><p><strong>Further Cases:</strong></p><p>Shai-Hulud 2.0 impacted over <strong>700 npm packages</strong> and <strong>25,000 GitHub repositories</strong>. Given that underlying libraries were downloaded hundreds of millions of times per week, thousands of DApp frontends unknowingly became vectors for fund theft globally.</p><p>The threat of software supply chain attacks is not limited to mass propagation; <strong>targeted attacks on individual projects</strong> also require vigilance. On <strong>December 25</strong>, Trust Wallet browser wallet version 2.68.0 was injected with a malicious backdoor, affecting numerous users and causing losses exceeding <strong>$7 million</strong>.</p><h3>3. AI Deepfake &amp; Social Engineering: Systemic Collapse of Digital Identity Trust</h3><p>The rapid development of AI has shifted social engineering attacks from static scams to <strong>dynamic, personalized interactions</strong>. In Web3 social engineering attacks, over <strong>56%</strong> involved AI deepfake technology, and losses due to AI deepfake combined with social engineering <strong>grew by 3000%</strong> in 2025.</p><p><strong>Mechanism:</strong></p><p>By leveraging large AI models, attackers can generate real-time <strong>fake video/audio</strong> and combine this with big data analysis of victims’ social footprints to conduct <strong>spear-phishing attacks</strong>. The core is simulating human interaction behavior: AI agents can engage in extended conversations, persuading users to install malware, share private keys, or authorize transactions.</p><p><strong>AI Deepfake Attacks Typically Follow Four Stages:</strong></p><ol><li><strong>Intelligence Gathering:</strong> Collect victim information from X, Telegram, Discord, and select targets.</li><li><strong>Fake Generation:</strong> Use AI tools to create video/audio deepfakes.</li><li><strong>Interactive Inducement:</strong> Build trust via video calls or voice chats, persuading users to install malware, share private keys, or perform transactions.</li><li><strong>Execution &amp; Escape:</strong> After asset transfer, traces are erased. Attack costs are significantly lower, and success rates increase by over 50%.</li></ol><p><strong>Further Cases:</strong></p><p>2025 saw a massive surge of AI deepfake and social engineering attacks in Web3. Cases are numerous, with notable examples including the <strong>UXLINK incident</strong> and the <strong>Venus high-net-worth phishing attack</strong>. More details:</p><ul><li><a href="https://x.com/UXLINKofficial/status/1998228868365529555">https://x.com/UXLINKofficial/status/1998228868365529555</a></li><li><a href="https://x.com/VenusProtocol/status/1962908345188450437?referrer=grok.com">https://x.com/VenusProtocol/status/1962908345188450437?referrer=grok.com</a></li></ul><h3>4. AI-Powered Smart Contract Vulnerability Discovery: The Arms Race of Automated Offense and Defense</h3><p>In 2025, <strong>AI-driven smart contract vulnerability discovery</strong> demonstrated the ability to independently identify and exploit smart contract flaws, accelerating a fundamental shift in the traditional paradigm of contract auditing and exploitation.</p><p><strong>Mechanism:</strong></p><p>AI models such as <strong>Claude Sonnet 4.5+</strong> and <strong>GPT-5+</strong>, trained on massive volumes of Solidity code, simulate execution paths to identify syntax errors, logical flaws, and other smart contract vulnerabilities. These models can then automatically generate exploit scripts, enabling <strong>fully automated vulnerability discovery and exploitation</strong>.</p><p><strong>AI-Based Smart Contract Vulnerability Discovery Typically Involves Three Stages:</strong></p><ol><li><strong>Vulnerability Scanning:</strong> AI analyzes contract code to identify zero-day vulnerabilities.</li><li><strong>Simulated Validation:</strong> AI evaluates and validates the exploitability of discovered vulnerabilities.</li><li><strong>Exploit Execution:</strong> AI generates attack code to carry out automated exploitation.</li></ol><p><strong>Further Cases:</strong></p><p>In 2025, long-running smart contracts such as <strong>Balancer</strong> and <strong>Yearn</strong>, despite years of secure operation, continued to have new vulnerabilities uncovered. This highlights the rapidly increasing capability of attackers to leverage AI for discovering <strong>deep, non-obvious smart contract vulnerabilities</strong>.</p><p>More details:</p><ul><li><a href="https://x.com/GoPlusSecurity/status/1985371222230884799">https://x.com/GoPlusSecurity/status/1985371222230884799</a></li><li><a href="https://x.com/GoPlusZH/status/1985360069031289085">https://x.com/GoPlusZH/status/1985360069031289085</a></li><li><a href="https://x.com/GoPlusSecurity/status/1995383979403518310">https://x.com/GoPlusSecurity/status/1995383979403518310</a></li><li><a href="https://x.com/GoPlusZH/status/1995477345654985181">https://x.com/GoPlusZH/status/1995477345654985181</a></li></ul><h3>Chapter 4: GoPlus Security Recommendations</h3><p>GoPlus believes that the future of Web3 security is no longer just about <strong>“catching bad actors” or “finding vulnerabilities”</strong>, but about providing <strong>ubiquitous, infrastructure-level (chain-native) and intelligent (AI-powered) protection</strong> for users’ assets — protection that users barely notice, yet feel completely secure.</p><p><strong>For Users:</strong></p><ul><li>Always maintain the habit of <strong>“Don’t click, don’t sign, don’t install, don’t transfer.”</strong> Do not click unknown links, sign unclear transactions, install untrusted software, or transfer funds to unfamiliar addresses.</li><li>Download wallet apps <strong>only from official channels</strong> and cross-verify via social media accounts, DefiLlama, CoinGecko, CMC, etc.</li><li>Install security plugins and stay alert for phishing links claiming “airdrops” or “prizes.”</li><li>Always <strong>verify signature contents and transaction intent</strong>; refuse blind signing.</li><li>When purchasing tokens, verify the <strong>token contract address</strong> and check token risk.</li><li>Do not store large amounts in hot wallets; buy cold wallets through official channels.</li><li>Keep private keys/mnemonics secure: handwrite multiple copies, store offline, protect from fire, moisture, and carelessness, <strong>never connect to the internet, screenshot, or upload to cloud</strong>. Anyone asking for your private key/mnemonic is <strong>100% a scammer</strong>.</li></ul><p><strong>For Projects and Chain Ecosystems:</strong></p><p>Projects and chains need to build a <strong>system-level security framework</strong> that goes beyond single-point security, covering the <strong>entire lifecycle from development to user interaction.</strong></p><ul><li><strong>Development Stage — Security Left Shift:</strong> Use secure contract templates to reduce vulnerabilities from the start; integrate AI auditing tools for continuous, automated vulnerability screening before deployment, addressing threats from AI-driven exploit discovery.</li><li><strong>Operation Stage — Full-Cycle Monitoring:</strong> Integrate security APIs to provide users with token detection, malicious address databases, authorization risk assessment, and transaction simulation; implement transparent lockups using multisig + asset locking tools to demonstrate liquidity security and increase trust.</li><li><strong>Response Stage — Rapid Emergency Mechanism:</strong> Collaborate with security teams and pre-establish communication with security service providers to respond quickly to incidents, minimizing losses.</li><li><strong>Native Security:</strong> Upgrade security from the <strong>“application layer”</strong> to the <strong>“protocol and base layers”</strong>, creating proactive security capabilities.</li></ul><p><strong>Security is the essential path for Web3 to achieve mass adoption. 2025 security data and technological trends indicate that security is no longer an after-the-fact cost — it is a fundamental infrastructure that determines whether Web3 can go mainstream.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1deb46dac16c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hyperliquid Buyback, Burn, and Staking Mechanism Research Report]]></title>
            <link>https://goplussecurity.medium.com/hyperliquid-buyback-burn-and-staking-mechanism-research-report-72e0e1765fd9?source=rss-176f1d291bf0------2</link>
            <guid isPermaLink="false">https://medium.com/p/72e0e1765fd9</guid>
            <category><![CDATA[goplus]]></category>
            <category><![CDATA[hyperliquid]]></category>
            <category><![CDATA[security]]></category>
            <dc:creator><![CDATA[GoPlus Security]]></dc:creator>
            <pubDate>Fri, 14 Nov 2025 10:44:44 GMT</pubDate>
            <atom:updated>2025-11-14T10:44:44.435Z</atom:updated>
            <content:encoded><![CDATA[<p>In-Depth Analysis of HYPE Token Economic Model</p><p>November 2025</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j3WRi8ia_EaEEzaB9OQu0A.png" /></figure><p><strong>Executive Summary</strong></p><p>Hyperliquid, as an innovative decentralized derivatives trading platform, has established a unique token economic model. This report provides an in-depth analysis of the three core mechanisms of the HYPE token: Buyback, Burn, and Staking, and combines the latest on-chain data to comprehensively evaluate the impact of these mechanisms on token value and ecosystem development.</p><p><strong>Key Findings:</strong></p><p>· Hyperliquid currently allocates 99% of trading fees to buy back HYPE tokens, establishing a powerful value accumulation mechanism, though not all bought-back HYPE is burned.</p><p>· The Assistance Fund conducts real-time daily buybacks.</p><p>· Based on the official announcement from January 19, 2024, the Future Fund may not only buy back $HYPE.</p><p>· Currently, approximately 41.8% of HYPE has been staked.</p><p>· The HYPE portion of HYPE-USDC spot trading fees is directly burned.</p><h3>1. Buyback Mechanism</h3><h4>1.1 Assistance Fund Overview</h4><p>The core of Hyperliquid’s buyback mechanism is the Assistance Fund (AF), an on-chain smart contract address responsible for executing automated buyback operations of protocol revenue. The establishment of the Assistance Fund embodies Hyperliquid’s ‘community-first’ philosophy, ensuring that the value generated by the protocol is directly returned to token holders.</p><p>It’s worth noting that although the Assistance Fund currently only buys back HYPE, according to official information released on January 19, 2024, the Assistance Fund may also buy back other tokens in the future.</p><p>Aid Fund Address: 0xfefefefefefefefefefefefefefefefefefefe</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pHZds8QXkIo2WnEISgWT1Q.png" /><figcaption>Figure 1: Hyperliquid Assistance Fund Official Announcement (January 19, 2024)</figcaption></figure><h4>1.2 Source of Buyback Funds</h4><p>Hyperliquid’s buyback mechanism has three major revenue sources. According to the official announcement on August 26, 2025, 99% of HLP fees automatically flow into the Assistance Fund for HYPE buybacks.</p><p>· <strong>Perpetual Contract Trading Fees</strong>: This is the primary revenue source.</p><p>· <strong>HIP-1 Auction Fees</strong>: Fees generated from new token listing auctions, with 100% used for HYPE buybacks.</p><p>· <strong>Spot Trading Fees</strong>: The USDC portion of trading fees is used for buybacks, while the HYPE portion is directly burned.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6G_HDTLgVknhsyKtxLBazQ.png" /><figcaption>Figure 2: Fee Distribution Mechanism Update Announcement (August 26, 2024)</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ViVAgSo3CNSud6e6K4UDsQ.png" /><figcaption>Figure 3: Hyperliquid Official Documentation — Fees Fully Returned to the Community</figcaption></figure><h4>1.3 How the Buyback Mechanism Works</h4><p>Unlike traditional dividend models, Hyperliquid employs a buyback mechanism that uses protocol revenue to purchase HYPE tokens from the secondary market. This design has several key features:</p><p>· <strong>Automated Execution</strong>: The Assistance Fund uses accumulated USDC to buy back HYPE in real-time, requiring no manual intervention.</p><p>· <strong>Open and Transparent</strong>: All buyback operations are executed on-chain, allowing anyone to track them.</p><p>· <strong>Continuous Buy Pressure</strong>: The greater the trading volume, the stronger the buyback force, creating a positive flywheel effect.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NNJKGU4Tor3Padr2KiTGUA.png" /><figcaption>Figure 4: Assistance Fund On-Chain Data — Holdings and Transaction Records</figcaption></figure><h4>1.4 Buyback Data Analysis</h4><p>According to the DUNE dashboard created by GoPlus:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/722/1*es1eWQLqac2R3yhBlz5png.png" /></figure><p>As of the last time this query was run, approximately 16.56M HYPE has been bought back, with a total value of approximately $570M USD.</p><h3>2、Burn Mechanism</h3><h4>2.1 Burn Mechanism Overview</h4><p>In addition to the buyback mechanism, Hyperliquid has also implemented a token burn mechanism, which permanently removes HYPE tokens from circulation to create scarcity. The burn mechanism complements the buyback mechanism, and together they work to enhance the value of the HYPE token.</p><h4>2.2 Burn Data</h4><p>According to the Dune dashboard created by GoPlus, as of the last time the query was run, a total of 471K HYPE tokens had been burned, which is significantly less than the total amount of HYPE repurchased by the Aid Fund.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*U5N6JrWqHkgfH01i5JWPLQ.png" /></figure><p>Hyperliquid’s buyback and burn mechanisms create a powerful synergistic effect. The buyback removes liquidity from the market, generating buying pressure, while the burn permanently reduces the supply. Together, they drive the value of the HYPE token from both the supply and demand sides.</p><h3>3、Staking Mechanism</h3><h4>3.1 Staking Mechanism Live</h4><p>On December 30, 2024, Hyperliquid officially launched the HYPE token staking feature. The staking mechanism is a key component of the consensus layer of the Hyperliquid L1 blockchain. It adopts a Delegated Proof of Stake (DPoS) model and ensures network security and efficient operation through the HyperBFT consensus algorithm.</p><h4>3.2 Staking Yield and Related Data</h4><p>Hyperliquid’s staking yield model is inspired by Ethereum’s design and adopts a dynamic adjustment mechanism that is inversely proportional to the square root of the total staked amount. This means that as more tokens are staked, the annualized yield decreases accordingly, ensuring the sustainability of the system.</p><p>Current Staking Data (November 2025):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ec-vcTRGUF74Qeb76MYBJQ.png" /></figure><p>According to the Dune dashboard created by GoPlus, as of the last time the query was run, approximately 418M HYPE tokens had been staked.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lFOdus5hg2QRm9WxYb34bQ.png" /></figure><h4>3.3 Staking Mechanism Features</h4><p>• <strong>Auto-Compounding:</strong> Staking rewards are distributed daily and automatically re-delegated to validators, achieving a compounding effect.<br>• <strong>Flexible Validator Selection:</strong> Users can delegate HYPE to any validator, with validator commission rates capped at 5%.<br>• <strong>Security Mechanism:</strong> A 7-day unlocking period prevents large-scale consensus attacks and protects network security.<br>• <strong>No Minimum Requirement:</strong> Any amount of HYPE can be staked, lowering the barrier to participation.</p><h4>3.4 Staking Tier System (Updated March 2025)</h4><p>In March 2025, Hyperliquid introduced a staking tier system, offering different trading fee discounts based on the amount staked. This mechanism further incentivizes users to hold and stake HYPE tokens long-term.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5r8uSZYz7AF4GgB8wCjFLw.png" /></figure><p>Other HyperLiquid-related data can be found in detail on the <strong>DUNE dashboard</strong> created by the GoPlus team：<a href="https://dune.com/goplus_security/hyperliquid-buyback-and-burn-and-staking-and-others">https://dune.com/goplus_security/hyperliquid-buyback-and-burn-and-staking-and-others</a></p><blockquote>End of Report<br>This report is for reference only and does not constitute investment advice. The cryptocurrency market carries high risks; invest with caution.</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72e0e1765fd9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>