<?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 Cloudsmith on Medium]]></title>
        <description><![CDATA[Stories by Cloudsmith on Medium]]></description>
        <link>https://medium.com/@cloudsmith?source=rss-2f200e7ca9d3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*i3X1xtMo8WgT0rwwlA37vg@2x.png</url>
            <title>Stories by Cloudsmith on Medium</title>
            <link>https://medium.com/@cloudsmith?source=rss-2f200e7ca9d3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 11:45:49 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@cloudsmith/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[A dependency firewall gives your scanner better inputs]]></title>
            <link>https://cloudsmith.medium.com/a-dependency-firewall-gives-your-scanner-better-inputs-359d245fab31?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/359d245fab31</guid>
            <category><![CDATA[dependency-firewall]]></category>
            <category><![CDATA[devsecops]]></category>
            <category><![CDATA[software-supply-chain]]></category>
            <category><![CDATA[supply-chain-security]]></category>
            <category><![CDATA[supply-chain-attack]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Tue, 14 Jul 2026 15:05:25 GMT</pubDate>
            <atom:updated>2026-07-14T15:05:25.612Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D-ocOOYuXM5EFBvUSjJ-XA.png" /></figure><p>Scanning for malicious packages at deployment catches issues already inside your environment. Scanning at the pull request catches what your CI pipeline already ran. If security controls don’t detect a malicious package until after ingestion, it has already had multiple opportunities to execute.</p><p>This is an architecture problem that requires an architectural answer. A dependency firewall is the answer. It’s a security posture that places guardrails between your development environment and public registries, evaluating what enters before any build runs.</p><h3>How malicious packages enter your build environment without inspection</h3><p>Build tools connecting directly to public registries like npm, PyPI, Maven Central, and Docker Hub, and packages arriving as requested is a common approach to open source software consumption. When scanning runs at the pull request, in the pipeline, or at deployment, the package is already inside your environment.</p><p>Software composition analysis (<a href="https://en.wikipedia.org/wiki/Software_composition_analysis"><strong>SCA</strong></a>) tools identify known vulnerabilities in packages already present in a codebase. That’s genuinely useful. But even with the best scanners, an architectural gap exists between ingestion and the location of the first scanner. Establishing a control point in that gap allows teams to better govern the packages that enter their environments.</p><p><strong>Two patterns produce this gap consistently.</strong></p><ul><li>The first is when no artifact management exists at all, and developers pull directly from public registries. Teams can’t implement security controls because the layer required isn’t in place.</li><li>The second, and it’s more common than most organizations realize, is when an existing artifact management tool handles the storage of software artifacts but enforces no security controls on their consumption (for example, AWS CodeArtifact or GitHub).</li></ul><p><strong>A third pattern is emerging due to the use of AI coding tools.</strong></p><ul><li>These agents generate code that expands the attack surface faster than most security programs have adjusted for. Tools like Cursor, AI-assisted editors, and LLM-powered agents recommend and install packages as a normal part of their workflow. Teams need governance that can match the pace of AI-powered development.</li></ul><p>Developers tend to accept those recommendations without the scrutiny they’d apply to a manual dependency addition. The packages arrive through the same registries and the same pipelines as they would with manually-written code, but the human review step between “<em>the model suggested it</em>” and “<em>it’s in the project</em>” is often gone.</p><h3>Modern attacks execute before the scanner runs</h3><p>The latest supply chain compromises exploit trusted open source packages. The <a href="https://cloudsmith.com/blog/axios-npm-attack-response">Axios</a>, <a href="https://cloudsmith.com/blog/how-cloudsmith-can-protect-against-the-litellm-attack">LiteLLM</a>, and the <a href="https://cloudsmith.com/blog/shai-hulud-the-second-coming-what-you-need-to-know-and-do-now">Shai-Hulud</a> attacks succeeded because they targeted specific packages developers already trusted and executed payloads at install time — before any scanner ran.</p><p>Cloudsmith’s analysis of 2025 OSV.dev data, comparing when malicious versions were published to npm against when OSV flagged them, shows that <strong>99 percent of npm package compromise attacks are identified within 72 hours of publication</strong>. A package scanned within that exposure window looks clean because no malware identifier exists yet. By the time the community flags the threat, pipelines are already running it.</p><h3>Scans are snapshots — threat intelligence isn’t</h3><p>There’s a second failure mode related to scanning that plays out more slowly. A package enters your environment, the scanner runs, nothing matches, and it’s marked clean. That evaluation is correct at that moment because no known vulnerability exists. Then six months later, a researcher discloses a critical CVE. Nothing triggers a re-evaluation, so the original point-in-time scan remains the most recent one and the package is still in production.</p><p>An example of this is the <a href="https://osv.dev/vulnerability/CVE-2021-44228"><strong>Log4Shell</strong></a> vulnerability in Apache Log4j. That package existed in production codebases for years before threat intelligence identified the vulnerability. The problem in cases like this is when you treat a one-time evaluation at the boundary as permanent clearance.</p><p>New advisories arrive constantly for packages that evaluate as clean at ingest. A scanner that runs once, when a package first arrives, doesn’t protect against what threat researchers find next month or next year. The scan result ages out the moment new intelligence contradicts it.</p><p>That said, scanning remains essential. But the fluid nature of supply chain threats, packages that evaluate as clean at ingest and become risks as new intelligence emerges, reveals a gap that point-in-time evaluation can’t close on its own. Addressing it means supplementing scanning with controls that operate earlier in the process, before packages enter an environment where they can execute.</p><h3>The control point that doesn’t depend on developer behavior</h3><p>A dependency firewall addresses both failure modes described above by changing where and when evaluation happens. The first is early detection at the boundary: identifying threatening or malicious dependencies before they enter your environment, rather than after they had the chance to execute. The second is ongoing assessment: unlike a scanner that runs once and treats the result as permanent, this posture evaluates continuously as new threat intelligence arrives. A package that cleared evaluation at ingest doesn’t remain cleared by default; it remains cleared only if current intelligence still supports that conclusion.</p><p>Developers start their work with packages that clear those guardrails, meaning the inputs to their work are more likely to be secure from the start. Vulnerabilities, threats, and license flags caught at the boundary save considerable rework compared to discovering the same issues in a PR review, a pipeline scan, or a post-incident forensic report.</p><p>For the vast majority of packages that clear evaluation, the process is entirely transparent and frictionless to developers. When friction occurs, like disabled scanner plugins, skipped linting rules, and manual approval workflows, developers create workarounds. Developers don’t need to bypass a control they don’t interact with. Compliance that relies on structural enforcement is more reliable than manual, behavior-based approaches.</p><p>For security and DevSecOps teams, the value is consistency at scale. Every developer, every AI agent, every pipeline, and every team in the organization hits the same guardrails. Policy applies universally and automatically, closing the gap between what policy says and what enforcement carries out.</p><p>A dependency firewall posture doesn’t replace the scanning you already have. It moves the trust decision to the ingestion boundary, so your team uses safe, evaluated packages before anyone writes a line of code.</p><p><em>For a practical guide to building a dependency firewall, covering controlled ingestion, continuous policy evaluation, and the OPA-backed policy engine, read </em><a href="https://cloudsmith.com/blog/how-to-use-cloudsmith-as-a-dependency-firewall"><em>How to use Cloudsmith as a dependency firewall</em></a><em>.</em></p><p>Or <a href="https://cloudsmith.com/book-a-demo"><em>book a demo</em></a><em> to see it in action.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=359d245fab31" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What is the SLSA framework? A guide to SLSA levels and supply chain security]]></title>
            <link>https://cloudsmith.medium.com/what-is-the-slsa-framework-a-guide-to-slsa-levels-and-supply-chain-security-7ea81d0bc438?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ea81d0bc438</guid>
            <category><![CDATA[slsa-levels]]></category>
            <category><![CDATA[devsecops]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[supply-chain-security]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Tue, 16 Jun 2026 12:52:49 GMT</pubDate>
            <atom:updated>2026-06-16T12:52:49.062Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F6mNkbnBb85P8X2QGLpxkg.png" /><figcaption>A guide to SLSA levels and supply chain security</figcaption></figure><p>Would you be able to prove how your software was built — from what source, on what infrastructure, under what controls — if a customer asked? The Supply-chain Levels for Software Artifacts (SLSA, pronounced “salsa”) framework is a structured answer to that problem.</p><h3>What is SLSA?</h3><p>SLSA is a security framework built around two fundamental questions: can you prove what built your software, and can you prove what source it came from? It answers both through a series of levels, each adding stronger guarantees around provenance, signing, and build or source controls.</p><p>SLSA is a technical standard for producing verifiable evidence that your software is what you say it is, not a compliance checkbox.</p><h3>History of SLSA</h3><p>A cross-industry consortium led by Google, in collaboration with the Open Source Security Foundation (OpenSSF) and the Linux Foundation, introduced SLSA in 2021. They built it in direct response to high-profile software supply chain attacks that exposed critical infrastructure vulnerabilities. The SolarWinds compromise, Codecov Bash Uploader breach, and Linux hypocrite commits all shared a common pattern: attackers tampered with build processes or injected malicious code into trusted software without detection.</p><p>The consortium designed SLSA to establish a standardized framework for securing the software supply chain based on real-world threats and security gaps. Since its initial release, SLSA has evolved significantly, particularly with the formalization of its track-based level structure.</p><p>As a vendor-neutral framework, it is an increasingly common benchmark for software supply chain security compliance, mapping closely to requirements in regulations like the EU Cyber Resilience Act (CRA), the Digital Operational Resilience Act (DORA), and the Federal Risk and Authorization Management Program (FedRAMP).</p><h3>Progressive security with SLSA</h3><p>SLSA is progressive in that it defines a series of levels of increasing supply chain security guarantees. Version 0.1 defined four conformance levels across three domains: build, source, and dependencies.</p><p>Version 1.0 revised that approach, splitting evaluation into separate tracks for each domain. Teams can measure and advance progress on one security dimension without being blocked by another. The current version, SLSA 1.2, formalizes both the build track and the source track, which together address the integrity of how software is built and where it originates.</p><h3>SLSA levels explained: The build track</h3><p>The build track focuses on the provenance of an artifact, the verifiable evidence of how, where, and by whom it was built. The build track has four levels:</p><ul><li><strong>SLSA Build Level 0 (“None”): </strong>No security controls.</li><li><strong>SLSA Build Level 1 (“Provenance exists”): </strong>Provenance only, describing how the package was built. The focus is on the manual detection of mistakes and having documentation in place.</li><li><strong>SLSA Build Level 2 (“Hosted build platform” ): </strong>The host platform signs and generates the provenance. The focus is on detecting and preventing tampering <em>after</em> the build.</li><li><strong>SLSA Build Level 3 (“Hardened builds”): </strong>A hardened, isolated build platform. This level focuses on detecting and preventing tampering <em>during</em> the build process.</li></ul><h3>SLSA Build Track Level 1: Laying the foundations</h3><p>SLSA Build Level 1 requires builds to produce provenance metadata, which includes details on what was built, how, by whom, and on what platform. This doesn’t guard against tampering, but it introduces transparency that is frequently lacking today. With provenance in place, you can reduce the risk of release errors and give your teams a working inventory of software components and build platforms.</p><p>You can configure most CI/CD systems to generate provenance in standard formats such as SPDX or CycloneDX. This improves accountability and visibility without heavy overhead, making it the practical starting point for SLSA compliance.</p><p>Cloud-native systems support Level 1 by automatically storing and serving artifacts with attached metadata and provenance, providing the required visibility into what was built and how.</p><h3>SLSA Build Track Level 2: Establishing trust with a hosted build platform</h3><p>Level 2 advances software supply chain security by requiring builds to run on trusted, dedicated infrastructure platforms like Google Cloud Build or GitHub Actions rather than a developer’s local machine. This deters casual tampering and mitigates risk from insider threats or attackers who face legal or financial consequences. The build platform digitally signs the provenance and ties it to the infrastructure, so artifacts are harder to fake or manipulate without a traceable, intentional attack.</p><p>To meet Level 2, the organization must use a hosted build system that automatically generates and signs provenance. Downstream consumers must also verify the authenticity of that provenance. Level 2 pushes teams toward more secure, auditable build platforms and strengthens release integrity at scale.</p><p>Cloud-native artifact management platforms offer hosted build environments that generate signed provenance and ensure artifacts are stored securely, verified on upload, and distributed globally. Integrations with CI/CD tools like CircleCI, GitHub Actions, and Google Cloud Build are straightforward.</p><h3>SLSA Build Track Level 3: Hardening builds against tamper resistance</h3><p>Level 3 runs the build process in a hardened, isolated environment, like a dedicated container or virtual machine, where no other build process can interfere and signing material is strictly protected. Provenance cannot be forged or bypassed without a serious exploit, which blocks the vast majority of adversaries, including those operating with compromised credentials.</p><p>Level 3 requires a secure build environment with strict process isolation and protected signing material. A cloud-native artifact management platform supports this by enforcing build isolation, working with established hardened CI systems, and securely storing signed provenance.</p><h3>SLSA levels explained: The source track</h3><p>The build track tells you how an artifact was built. The source track tells you where the code came from and whether the process that produced it can be trusted. SLSA 1.2 formalizes the source track as a standalone progression with four levels, each adding stronger guarantees about version control, history integrity, enforced controls, and code review.</p><p>Together, the build and source tracks address the integrity picture end to end: was this code reviewed and controlled before it was built, and was the build itself trustworthy?</p><p><strong>Where does your organization stand on the SLSA scale?</strong> Find out which SLSA level you’re operating at and how that maps to frameworks like CRA, DORA, and FedRAMP. <a href="https://cloudsmith.com/resources/artifact-security-maturity-tool">Take the free Security Maturity Assessment</a> to benchmark your software supply chain in a few minutes.</p><h3>Cloud-native and SLSA compliance</h3><p><a href="https://cloudsmith.com/resources/accelerate-slsa-compliance">Achieving SLSA compliance</a> requires strong provenance, integrity guarantees, and a secure, verifiable software supply chain. While cloud-native registries inherently align with these requirements, self-managed or on-premises registries introduce risks that may compromise compliance — especially at SLSA Build Level 2, which mandates a separation between CI/CD and artifact distribution.</p><p>Cloud-native artifact management platforms offer several advantages that directly support SLSA compliance:</p><ul><li><strong>Built-in provenance and immutability</strong>: Cloud-native registries natively support provenance metadata and artifact immutability, ensuring tamper-resistant storage.</li><li><strong>Automated signing and access controls</strong>: Enforced signing and fine-grained access policies prevent unauthorized modifications, while audit logs provide verifiable tracking of all changes.</li><li><strong>End-to-end traceability</strong>: Visibility across builds and releases enables transparent tracking of software artifacts, a cornerstone of SLSA’s integrity and provenance standards.</li><li><strong>Scalability and global reliability</strong>: Enterprise-grade cloud-native platforms dynamically scale to worldwide demand, ensuring reliable artifact distribution while maintaining security and compliance.</li></ul><p>Choosing a cloud-native registry directly simplifies and strengthens SLSA compliance by leveraging built-in capabilities around reproducibility, integrity, and security.</p><h3>Practical challenges for non-cloud solutions</h3><p>Attempting to achieve SLSA compliance with on-premises or bespoke solutions introduces significant challenges that compound as you move up the levels:</p><ul><li><strong>Increased risk of human error:</strong> Manual artifact handling raises the chances of accidental modifications or tampering.</li><li><strong>Security and compliance gaps</strong>: Self-hosted registries lack built-in immutability guarantees, increasing the risk of provenance loss or unauthorized changes.</li><li><strong>Scalability and consistency issues</strong>: Managing distributed storage across multiple regions is complex, often leading to inconsistent security postures and availability risks.</li><li><strong>Lack of automated provenance</strong>: Without native support for signed metadata and verification, guaranteeing artifact authenticity becomes harder and error-prone.</li><li><strong>Poor separation of concerns</strong>: On-prem solutions can blur the boundary between developer access and infrastructure control, potentially exposing critical build processes to risk.</li></ul><p>That last point is the most consequential failing because it directly undermines SLSA Build Level 2’s requirement for separation between CI/CD and artifact distribution. A hosted environment solves what the framework treats as a minimum requirement.</p><h3>FAQs: SLSA framework and SLSA levels</h3><ol><li><strong>What is the SLSA framework?</strong></li></ol><p>SLSA (Supply-chain Levels for Software Artifacts) is an industry-backed software supply chain security framework that defines progressive standards for verifying how software is built, sourced, and maintained. A cross-industry consortium led by Google, in collaboration with OpenSSF and the Linux Foundation, created it in 2021.</p><p><strong>2. What are the SLSA levels?</strong></p><p>SLSA 1.2 defines two tracks, each with multiple levels. The build track has four levels (0–3), focusing on provenance and build integrity. The source track has four levels (1–4), focusing on version control, history integrity, enforced technical controls, and two-party review.</p><p><strong>3. What is SLSA Level 1?</strong></p><p>SLSA Build Level 1 requires builds to produce provenance metadata describing what was built, how, and by whom. It’s the easiest level to implement because it focuses on transparency rather than tamper prevention.</p><p><strong>4. What is the difference between SLSA Build Level 2 and Level 3?</strong></p><p>Build Level 2 requires builds to run on a hosted platform that automatically signs and generates provenance, focused on detecting tampering <em>after</em> the build. Level 3 requires a hardened, isolated build environment that prevents tampering <em>during</em> the build itself.</p><p><strong>5. Is SLSA compliance mandatory?</strong></p><p>No, SLSA is a voluntary framework. However, it maps closely to requirements in regulations like the CRA, DORA, and FedRAMP, making it an increasingly common benchmark for security compliance and software procurement requirements.</p><p><strong>6. How do I know what SLSA level my organization is at?</strong></p><p>Assess your build, source, and provenance practices against SLSA’s criteria for each level. Cloudsmith’s<a href="https://cloudsmith.com/resources/artifact-security-maturity-tool"> Security Maturity Assessment</a> walks you through this and benchmarks your results against CRA, DORA, and FedRAMP signals.</p><p><strong>7. Is SLSA the same as an SBOM?</strong></p><p>No. An SBOM (Software Bill of Materials) is an inventory of a software’s components, while SLSA is a framework for securing the build and provenance process itself. SLSA provenance can include or reference an SBOM, but they serve different purposes.</p><p><strong>8. What is the SLSA source track?</strong></p><p>The SLSA source track is a set of four levels that define increasing trustworthiness in how source code is produced and controlled: from basic version control at Level 1, through immutable history and source provenance attestations at Level 2, to enforced technical controls at Level 3, and mandatory two-party review at Level 4. It was formalized in SLSA 1.2 alongside the build track.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ea81d0bc438" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The EU Cyber Resilience Act: What engineering teams need to do to be compliant]]></title>
            <link>https://cloudsmith.medium.com/the-eu-cyber-resilience-act-what-engineering-teams-need-to-do-to-be-compliant-86aaac6d7a55?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/86aaac6d7a55</guid>
            <category><![CDATA[eu-cyber-resilience-act]]></category>
            <category><![CDATA[cyber-resilience-act]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[compliance]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Thu, 11 Jun 2026 12:02:27 GMT</pubDate>
            <atom:updated>2026-06-11T12:02:27.847Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*3QLzwo-TZ42F-NZKTj8EUw.png" /><figcaption>The EU Cyber Resilience Act</figcaption></figure><p>The EU Cyber Resilience Act represents a meaningful shift in how software security is defined, enforced, and evidenced, with engineering teams at the center. With Cyber Resilience Act compliance now an active requirement and the September 11, 2026, vulnerability reporting deadline five months out, the path forward is clear: teams that build the right foundations now, SBOMs, automated vulnerability monitoring, and governed artifact pipelines, won’t just achieve CRA compliance in 2026. They’ll ship software their customers can trust at a higher standard than before.</p><p>This isn’t a regulation that lives in a legal team’s inbox. The requirements land squarely on your codebase, CI/CD pipeline, artifact repository, and how you track and respond to vulnerabilities.</p><p>This blog is for the architects, DevOps leads, and security engineers doing the actual work, not just the “what” but the “how” and the “now”.</p><h3>What is the EU Cyber Resilience Act (CRA)?</h3><p>The EU Cyber Resilience Act (Regulation EU 2024/2847) is the first EU-wide regulation that places mandatory cybersecurity requirements on digital products throughout their entire lifecycle, from design and development all the way through deployment and end-of-life.</p><p>Before the CRA, cybersecurity obligations in the EU were fragmented. Regulations like NIS2 and GDPR addressed data protection and critical infrastructure, but nothing set a consistent product-level security baseline for software and hardware. The CRA fills that gap. It establishes a common framework for how companies must design, maintain, and support digital products, and it applies to anyone manufacturing, importing, or distributing products with digital elements into the EU market.</p><h3>The 2026 deadline: The 24-hour vulnerability reporting rule</h3><p>The most urgent date on your calendar is <strong>September 11, 2026</strong>. This is when <strong>Article 14</strong> takes effect. It requires manufacturers to report <strong>actively exploited vulnerabilities</strong> to the ENISA Single Reporting Platform (SRP) and national CSIRTs.</p><p>The reporting window is punishingly tight:</p><ol><li><strong>Early Warning (24 Hours):</strong> Within 24 hours of becoming aware of an actively exploited vulnerability or a severe incident, you must submit an “early warning” to ENISA.</li><li><strong>Full Notification (72 Hours):</strong> Within 72 hours, you must provide a detailed report including an initial assessment of the vulnerability’s impact and any corrective measures taken.</li><li><strong>Final Report (14 Days):</strong> A final report must be submitted no later than 14 days after a corrective measure (a patch or mitigation) is available.</li></ol><p><strong>The Reality Check:</strong> To comply by September, your team needs an automated <strong>vulnerability handling</strong> process that can triage alerts from your <strong>artifact repository and CI/CD pipeline</strong> in real-time.</p><h3>What CRA software requirements look like for engineering teams</h3><p>The CRA introduces 21 essential cybersecurity requirements for manufacturers. From an engineering perspective, four areas carry the most weight:</p><h3>1. Secure by design</h3><p>The CRA requires that companies build security into the product from the start, not bolted on after shipping. Concretely, that means products must ship with secure default configurations, minimize unnecessary attack surfaces, and include protections against common vulnerability classes. You can’t ship with default credentials or insecure-by-default settings and expect to remain compliant.</p><p>For DevOps teams, this means security needs to be part of how you design features and structure your release process, not a checkbox at the end of a sprint.</p><h3>2. Software Bill of Materials (SBOM)</h3><p>The CRA requires manufacturers to maintain an SBOM, a machine-readable inventory of every component in a product, including open-source dependencies, third-party libraries, and any other software elements. The SBOM needs to be accurate and kept up to date.</p><p>The dependency between SBOMs and the September 2026 reporting deadline is real: With full component-level visibility into your products, you can report on an actively exploited vulnerability quickly and with confidence, which is exactly what the September 2026 deadline demands. Practically speaking, SBOM readiness is required for the 2026 deadline even though the formal SBOM mandate is part of the 2027 requirements.</p><p>A repeatable <a href="https://cloudsmith.com/blog/how-to-generate-and-host-an-sbom">SBOM generation process</a> is the logical place to start.</p><h3>3. Vulnerability handling and reporting</h3><p>The CRA establishes structured requirements for how manufacturers handle vulnerabilities, not just the 24-hour reporting obligation but the ongoing process of tracking, triaging, patching, and disclosing them. This includes:</p><ul><li>A documented vulnerability management process covering your entire product portfolio</li><li>The ability to push security updates in a timely manner across the product’s supported lifecycle</li><li>A public policy for coordinated vulnerability disclosure</li><li>Documentation of how vulnerabilities were resolved and when</li></ul><p>This shifts vulnerability management from an ad hoc engineering practice to a compliance-critical, documented discipline.</p><h3>4. Technical documentation</h3><p>The CRA requires manufacturers to maintain a detailed technical documentation package that covers the product’s design, security risk assessment, SBOM, test results, and patch history. This documentation must be retained for 10 years and made available to market surveillance authorities on request.</p><p>For most engineering teams, this means treating documentation as a first-class artifact, something that you generate, version, and maintain alongside your code and releases.</p><h3>What is “secure by design” under the CRA?</h3><p>Secure by design isn’t a vague principle under the CRA; it’s a verifiable, documented requirement. Among the specific obligations, manufacturers must ensure that products:</p><ul><li>Protect the confidentiality and integrity of stored, transmitted, and processed data</li><li>Limit privileges and apply least-privilege access controls</li><li>Ship with security enabled by default, not just available as an option</li><li>Have the capability to receive and apply security updates throughout the supported lifecycle</li><li>Minimize the attack surface, including disabling unused interfaces and services</li></ul><h3>What are the penalties for CRA non-compliance?</h3><p>The EU has made it clear: the CRA is not a “paper tiger”. The penalties are tiered to match the severity of the oversight:</p><ul><li><strong>Tier 1 </strong>— Essential cybersecurity requirements and core manufacturer obligations (including the Article 14 vulnerability reporting obligations): fines of up to €15 million or 2.5% of total worldwide annual turnover, whichever is higher.</li><li><strong>Tier 2 — </strong>Other CRA obligations, including those applying to importers and distributors and conformity assessment requirements: fines of up to €10 million or 2% of worldwide annual turnover.</li><li><strong>Tier 3 — </strong>Providing false or misleading information to notified bodies or market surveillance authorities: fines of up to €5 million or 1% of worldwide turnover.</li></ul><p>For a company with €1 billion in revenue, 2.5% of global turnover is €25 million, well above the fixed ceiling. The fines are designed to sting regardless of company size.</p><p>Beyond financial penalties, market surveillance authorities can order product withdrawals, market bans, and product recalls. Losing access to the EU market is often more damaging than the fine itself.</p><h3>What engineering teams need to do to be compliant</h3><p>You don’t need to solve everything at once. You don’t need to solve everything at once. The teams in the best shape are simply the ones that started earliest. Here’s a practical breakdown:</p><ul><li><strong>Start with a product inventory: </strong>Map every product your organization ships to EU customers. For each one, determine whether it qualifies as a “product with digital elements” under the CRA, as nearly all commercial software does. Understanding your responsibilities under the CRA is the first step to scoping your compliance work.</li><li><strong>Build a repeatable SBOM generation process: </strong>Your SBOM is the foundation for almost everything else the CRA requires. It helps you monitor for vulnerable components, respond to the 24-hour reporting obligation, and meet the documentation requirements. Integrating SBOM generation into your build pipeline, so every release automatically produces an up-to-date, accurate SBOM, is the highest-leverage action most teams can take right now.</li></ul><p>An SBOM generation tool that integrates with your CI/CD pipeline and supports standard formats such as SPDX or CycloneDX is essential for CRA compliance. Manual SBOM processes won’t scale across a product portfolio.</p><ul><li><strong>Establish continuous dependency management and monitoring</strong></li></ul><p>Once you have <a href="http://cloudsmith.com/blog/how-to-analyze-an-sbom">SBOMs</a>, you need the ability to monitor them continuously. New vulnerabilities are disclosed daily, and the CRA’s reporting requirements apply to actively exploited vulnerabilities even in products you shipped years ago. Dependency management for CRA compliance means automated scanning against vulnerability databases, alert routing to the right teams, and a documented process for assessing whether a given vulnerability affects your products.</p><ul><li><strong>Evaluate your artifact repository for CRA readiness</strong></li></ul><p>Your artifact repository is the control point for everything your engineering team builds and ships. Artifact repository CRA compliance means having the ability to enforce policies on what components can enter your build pipeline, track provenance of every artifact, and generate the audit trails needed for technical documentation. If your current repository doesn’t support policy enforcement, vulnerability scanning, or SBOM integration, those are gaps worth addressing now.</p><ul><li><strong>Document your vulnerability handling process</strong></li></ul><p>Before September 2026, you need a written, operational process for detecting, triaging, and reporting actively <a href="https://cloudsmith.com/blog/how-to-manage-your-vulnerability-workflows-with-cloudsmith">exploited vulnerabilities</a>. This doesn’t need to be elaborate, but it does need to exist, be tested, and be understood by the people who will execute it under time pressure. Think of it like a fire drill for your incident response team.</p><ul><li><strong>Integrate security into your development lifecycle</strong></li></ul><p>The CRA’s secure-by-design requirements aren’t a one-time project. They’re a shift in how your team operates. Threat modeling, security testing, and secure configuration reviews need to become standard steps in your development process, with documented outputs that regulators can review if needed.</p><h3>The role of Cloudsmith in your CRA strategy</h3><p>Cloudsmith is a cloud-native artifact management platform that sits at the center of your <a href="https://cloudsmith.com/product/software-supply-chain-security">software supply chain</a>, controlling what goes in and what comes out of your build pipelines. Cloudsmith provides the infrastructure needed to automate these requirements:</p><ul><li><strong>Artifact repository CRA Compliance</strong></li></ul><p>Cloudsmith acts as a secure buffer. It allows you to ingest third-party dependencies, scan them, and automatically generate the necessary SBOMs. By centralizing your artifacts, you create a single audit trail for regulators.</p><ul><li><strong>Automated SBOM generation and hosting</strong></li></ul><p>Cloudsmith doesn’t just scan; it generates and hosts SBOMs. When a regulator or customer requests the SBOM for a specific 2025 version, Cloudsmith provides the instant, version-linked retrieval required by the CRA’s documentation mandates.</p><ul><li><strong>Policy enforcement</strong></li></ul><p>With Cloudsmith’s Policy Engine, your team defines the security baseline, and the platform enforces it. Set a quarantine policy for artifacts with unpatched vulnerabilities or missing provenance data, and nothing that fails that bar enters your pipeline.</p><h3>Conclusion: Compliance is the new quality standard</h3><p>The <strong>EU Cyber Resilience Act</strong> is more than just a regulatory hurdle; it is a fundamental change in how we define “production-ready” software. By September 2026, the ability to see into your <a href="https://cloudsmith.com/blog/the-2026-guide-to-software-supply-chain-security-from-static-sboms-to-agentic-governance">software supply chain</a> and report exploits in real time will be the price of entry into the European market.</p><p>Engineering teams that embrace <strong>Cyber Resilience Act compliance</strong> early, by automating SBOMs, securing their artifact repositories, and operationalizing vulnerability management, will not only avoid massive fines but also build greater trust with their global customer base.</p><p><strong>Cloudsmith gives engineering teams everything they need to meet the September 2026 deadline and build a stronger software supply chain in the process.</strong></p><h3>Frequently asked questions: EU Cyber Resilience Act</h3><ul><li><strong>Who does the Cyber Resilience Act apply to?</strong></li></ul><p>Any organization manufacturing, importing, or distributing products with digital elements in the EU, including commercial software, SaaS, IoT devices, and connected hardware. Note that CRA open source software exemptions only apply to non-commercial projects; commercial use puts you in scope.</p><ul><li><strong>What is the difference between the 2026 and 2027 CRA deadlines?</strong></li></ul><p>On September 11, 2026, vulnerability reporting obligations go live; manufacturers must report actively exploited vulnerabilities to ENISA within 24 hours, even for products already on the market. December 11, 2027, is the full compliance deadline, covering all remaining CRA software requirements, including SBOMs, secure-by-design, CE marking, and technical documentation for any new product placed on the EU market.</p><ul><li><strong>Do I need an SBOM to comply with CRA?</strong></li></ul><p>Yes. The CRA requires manufacturers to produce and maintain a software bill of materials for every product. Even though the formal SBOM mandate is part of the December 2027 deadline, you effectively need it before September 2026, because accurate component-level visibility is a prerequisite for meeting the 24-hour vulnerability reporting obligation. An SBOM generation tool integrated into your build pipeline is the most practical starting point.</p><ul><li><strong>How should DevOps teams prepare for the Cyber Resilience Act?</strong></li></ul><p>Start by integrating SBOM generation into your CI/CD pipeline and establishing continuous vulnerability monitoring across your product portfolio. From there, enforce artifact repository policies to prevent non-compliant components from entering your builds, and document your vulnerability-handling process before the September 2026 reporting deadline. The CRA effectively makes DevSecOps practices a legal requirement, not just a best practice.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=86aaac6d7a55" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Vibe coding vs. verified coding: Staying secure in the era of “prompt and pray.”]]></title>
            <link>https://cloudsmith.medium.com/vibe-coding-vs-verified-coding-staying-secure-in-the-era-of-prompt-and-pray-87073eb5ee6b?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/87073eb5ee6b</guid>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[artifact-management]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[devsecops]]></category>
            <category><![CDATA[ai-security]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Mon, 25 May 2026 10:16:13 GMT</pubDate>
            <atom:updated>2026-05-25T10:16:13.379Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Cjr_GsBOeLdp0-UCT5LTQ.png" /></figure><p><strong>TL;DR</strong></p><ul><li><strong>Vibe coding</strong>, the practice of shipping AI-generated code with minimal review, is now mainstream. It’s a distinct workflow from simply using AI to assist coding: 42% of developers report that AI writes at least half their codebase, but not all of them are skipping the review.</li><li>AI-generated code introduces supply chain risks that most CI/CD pipelines weren’t designed to catch: hallucinated packages, hardcoded secrets, and a 10x higher rate of security findings versus human-authored code.</li><li><strong>Verified coding</strong> is the practice of applying structured governance between generation and deployment, first enforcing policy at ingestion, then signing artifacts for traceability, and finally verifying provenance at the registry before anything reaches production.</li><li>The speed of generation and security of delivery don’t have to be in conflict, but keeping them aligned requires deliberate tooling and configuration choices at each layer of your pipeline.</li></ul><h3>What is vibe coding, and why is it a supply chain problem?</h3><p>Vibe coding is the practice of describing software intent in natural language, letting an AI assistant generate the implementation, and shipping without deep line-by-line review. <a href="https://x.com/karpathy/status/1886192184808149383"><strong>Andrej Karpathy</strong></a> coined the term in early 2025, but the workflow began quietly spreading through engineering teams long before it had a name.</p><p>It’s productive. It’s intuitive. And it introduces a class of risk that most organizations are still catching up to.</p><p>According to Cloudsmith’s 2026 <a href="https://cloudsmith.com/campaigns/2026-artifact-management-report"><strong>Artifact Management Report</strong></a>, 42% of developers using AI report that at least half of their codebase is now AI-generated. That’s a structural shift in how software gets built. When the majority of your artifacts have a non-human author, the question of provenance becomes urgent in a way it simply wasn’t before.</p><p>The supply chain problem is that AI writes code <em>fast</em>, faster than review cycles, faster than policy gates, and faster than the governance frameworks most security teams have in place. Vibe coding, at scale, is essentially an uncontrolled dependency entering your pipeline at commit velocity.</p><h3>What does “verified coding” actually mean?</h3><p>Verified coding is the practice of treating AI-generated code with the same, or greater, scrutiny as any third-party dependency entering your software supply chain. It doesn’t mean abandoning AI tooling or slowing down development. It means applying governance at the artifact layer: enforcing policy at the point of ingestion, signing and attesting what gets promoted, and ensuring that the speed of generation doesn’t outpace the security of delivery.</p><p><strong>Put simply: vibe coding gets code written. Verified coding gets code shipped <em>safely</em>.</strong></p><p>The distinction matters because the two workflows touch different parts of the pipeline. AI tools operate at the generation layer, the IDE, the prompt, and the initial commit. Verified coding operates at the artifact layer, the registry, the promotion gate, and the deployment boundary. These aren’t competing approaches. They’re sequential ones. And most teams currently have only the first half in place.</p><h3>Why “prompt and pray” is a legitimate security risk</h3><p>The numbers here are worth sitting with.</p><p>In a landmark <a href="https://www.veracode.com/blog/genai-code-security-report/">study</a>, <strong>Veracode</strong> tested over 100 large language models (LLMs) on security-sensitive coding tasks and found that <strong>45% of AI-generated code introduces OWASP Top 10 vulnerabilities</strong>, a failure rate that hasn’t meaningfully improved across multiple test cycles from 2025 through early 2026. Java saw a <strong>72% failure rate</strong>, while JavaScript, Python, and C# hovered around 40%.</p><p>This isn’t a model quality problem that will resolve itself in the next release cycle. It’s a training data problem: models are trained on public repositories saturated with legacy patterns, deprecated APIs, and insecure-by-default examples. When a model generates a database query, it may reach for string concatenation over parameterized queries, not because it doesn’t know better, but because that pattern appears more frequently in its training set. The insecure output is, statistically, the expected output.</p><p>The supply chain risks compound from there. Recent research from <a href="https://apiiro.com/blog/4x-velocity-10x-vulnerabilities-ai-coding-assistants-are-shipping-more-risks/">Apiiro</a> found that repositories with high AI adoption saw a <strong>322% increase in privilege-escalation flaws</strong> and a <strong>153% spike in architectural design flaws.</strong> This isn’t just about typos; it’s about structural weaknesses, things like insecure authentication flows, over-permissioned roles, missing input validation, and architectural patterns that expose blast radius when something goes wrong.</p><p>While AI-assisted developers are shipping commits at <strong>three to four times the rate</strong> of their peers, they are introducing security findings at <strong>ten times the rate</strong>. We are effectively trading human keystrokes for an exponential explosion in security debt.</p><p><strong>More commits. More artifacts. More surface area. Less human review at each handoff.</strong></p><p>And then there’s <strong>slopsquatting</strong>, arguably the most novel risk in the vibe coding era. AI tools occasionally hallucinate package names that don’t exist. Attackers now register those predicted names on npm and PyPI in advance. No phishing. No credential theft. Just the predictable output of a language model, exploited as a supply chain attack vector. A developer runs npm install, the package resolves to the attacker’s version, and malicious code enters the build.</p><h3>How does verified coding work in practice?</h3><p>Verified coding isn’t a single tool or a checklist. It’s a posture that primarily operates at the artifact layer, not the code-review layer. Here’s what it looks like in practice:</p><ul><li><strong>Policy enforcement at ingestion. </strong>Before an AI-generated package or dependency enters your pipeline, it is subject to the same policy gates as anything else: vulnerability scanning, license validation, and, increasingly, provenance attestation. If you can’t establish where an artifact came from and what generated it, that uncertainty is itself a signal.</li><li><strong>Signing and attestation. </strong>Verified coding requires a chain of custody. Cryptographically signed artifacts with verifiable provenance mean that if something goes wrong downstream, you can establish exactly when and where an untrusted artifact entered the supply chain. This is the difference between a security incident and a security audit.</li><li><strong>Immutable promotion pipelines.</strong> AI tools accelerate the pace of development. Immutable artifact repositories, where what gets promoted to production is fixed, versioned, and auditable, provide the anchor point that keeps that pace from becoming a liability.</li><li><strong>Separation of generation from governance. </strong>The developer using Cursor or Copilot shouldn’t also be the one deciding what reaches production. Verified coding means applying a governance layer between generation and deployment, ideally automated, policy-driven, and integrated into the CI/CD pipeline rather than added afterward.</li></ul><h3>Where does the artifact registry fit in?</h3><p>Most security conversations about vibe coding focus on code review. Did a human look at this before it shipped? That’s a reasonable instinct, but it’s the wrong layer to rely on as your primary control.</p><p>By the time a developer is reviewing AI-generated code, the artifact may already be in a dependency. It may already be in a container image. Code review doesn’t catch slop squatting, doesn’t enforce provenance, and doesn’t scale to the volume that AI-assisted development produces.</p><p>The artifact repository is where the supply chain consolidates, and therefore where governance has the greatest leverage. A secure artifact management platform doesn’t care whether your code was written by a developer, a junior engineer, or a language model. It applies a consistent policy to everything that flows through it. That’s the right abstraction level for AI-era security.</p><p>This is the shift from reactive scanning to proactive governance, moving the security gate from “what did we catch after the fact?” to “what does policy permit us to allow in?” For teams generating code at the scale and speed that AI enables, that shift is the architectural decision that determines whether your supply chain is manageable or not.</p><h3>Vibe coding and verified coding are not in conflict</h3><p>It’s worth saying clearly: this isn’t an argument against AI-assisted development. The productivity gains are real, the adoption is irreversible, and the teams generating code with AI will generally outship the teams that aren’t.</p><p>The argument is about where governance sits. Vibe coding moves fast at the generation layer. Verified coding holds the line at the artifact layer. These are different concerns, operating at different points in the pipeline, and conflating them is how teams end up either over-restricting AI tooling or under-governing what it produces.</p><p>Prompt and pray is a workflow. Verified coding is a standard. The gap between the two is where breaches happen, not dramatically, not all at once, but incrementally, through unexamined artifacts accumulating quietly in production.</p><h3>About Cloudsmith</h3><p>Cloudsmith is a secure, universal artifact management platform built for modern software supply chains. We help engineering and security teams enforce policy, manage provenance, and govern the artifacts flowing through their pipelines, regardless of how those artifacts were generated.</p><p><strong>If your team is building at an AI scale,</strong><a href="https://cloudsmith.com/"><strong> explore Cloudsmith</strong></a><strong>.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=87073eb5ee6b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 red flags in your open source dependencies that put your supply chain at risk]]></title>
            <link>https://cloudsmith.medium.com/5-red-flags-in-your-open-source-dependencies-that-put-your-supply-chain-at-risk-d7435102bb48?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/d7435102bb48</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[dependency-management]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Mon, 11 May 2026 10:23:38 GMT</pubDate>
            <atom:updated>2026-05-11T10:23:38.820Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yo-g-sLQGMwiKl58mNYaAg.png" /><figcaption>Is your supply chain at risk?</figcaption></figure><p>Open source is one of the best things that has ever happened to software development.</p><p>In a matter of minutes, you can pull in battle-tested libraries, frameworks, and tools that would take months to build from scratch and ship faster because of it. The modern software stack is a genuine collective achievement, and your ability to move quickly depends on it.</p><p>But as your stack matures, so should your relationship with the packages inside it. The teams shipping the most reliable, secure software today are the ones who are intentional about how they manage it.</p><p>At <a href="https://cloudsmith.com/"><strong>Cloudsmith</strong></a>, we work with engineering teams across every stage of the maturity curve. And there are five signals that consistently separate teams with strong supply chain hygiene from teams that are one upstream change away from a very bad week.</p><h3>What is open-source dependency risk, and why does it matter now?</h3><blockquote><strong>Definition: </strong>Open source dependency risk refers to the security, integrity, and reliability threats that enter your software supply chain through third-party packages, libraries, modules, and components your codebase depends on but doesn’t own.</blockquote><p>Modern applications are built using hundreds of open-source components, most of which no one on your team ever reviewed directly. And attackers notice. Supply chain attacks have grown significantly year over year, targeting the dependency layer precisely because it sits outside the boundaries of most security programs.</p><p>The question isn’t whether your dependencies carry risk. They do. The question is whether you can see that risk before it becomes an incident.</p><h3>Red flag #1: Unmaintained packages with no clear successor</h3><p><strong>What to look for: </strong>The last commit was over 12 months ago, open issues are piling up unanswered, and the maintainer’s GitHub profile hasn’t been active in over two years.</p><p>Abandoned packages are an underestimated risk in open source. They don’t just stop receiving features; they stop receiving security patches. When a CVE drops affecting an unmaintained package, your options narrow fast: patch it yourself, fork it, or rip it out entirely.</p><p>But the deeper risk is account takeover. Dormant packages with high download counts are prime targets for attackers who approach burned-out maintainers, or simply wait for expired domains and lapsed NPM tokens, to inject malicious code into a trusted package name.</p><p><strong>What to do:</strong></p><ul><li>Audit your dependency tree for packages with no meaningful activity in 12+ months.</li><li>Check if the original repository links to a successor or archive notice.</li><li>If not, treat it as a liability worth addressing in your next sprint cycle, not your next incident retrospective.</li></ul><h3>Red flag #2: Overly permissive or mismatched licenses</h3><p><strong>What to look for: </strong>GPL, AGPL, or SSPL licenses mixed into a commercial codebase without legal review. License fields marked as UNKNOWN or left entirely blank.</p><p>This is a legal risk and a signal of supply chain integrity. Packages with inconsistent or missing license declarations are often poorly governed, hastily published, or potentially forked from another project without proper attribution.</p><p>For commercial teams, a single AGPL dependency buried in your dependency graph can legally require you to open-source your entire application. Most teams discover this during a compliance audit or an acquisition due diligence process, the worst possible times to find out you’ve been “poisoned” by a restrictive license.</p><p><strong>What to do:</strong></p><ul><li>Run a license compliance scan across your full dependency tree, including transitive dependencies.</li><li>Configure your artifact repository to automatically surface license violations at ingestion, before problematic packages reach your build pipeline.</li><li><a href="https://cloudsmith.com/product/enterprise-policy-manager">See how Cloudsmith handles license policy enforcement →</a></li></ul><h3>Red flag #3: Packages from personal accounts with no organizational oversight</h3><p><strong>What to look for: </strong>The package maintainer is a single individual with no organizational affiliation, no 2FA enforced on their registry account, and no documented ownership transfer policy.</p><p>This is a structural risk, not a personal one. Individual maintainers get burnt out, change jobs, lose access to old email accounts, or simply decide they’re done. When that happens to a package your platform depends on, you have no escalation path.</p><p>And the risk goes beyond abandonment. A single compromised credential on a personal account is all it takes for an attacker to push a malicious version of a widely trusted package, one that your CI system will pull automatically on the next build, with no human review in between. The packages most at risk aren’t obscure ones nobody uses; they’re the quiet, stable ones that everyone depends on and nobody thinks twice about.</p><p><strong>What to do:</strong></p><ul><li>For critical dependencies, verify whether the package has organizational backing or a documented governance model.</li><li>Consider proxying packages through a private artifact repository so you control exactly which versions enter your builds, regardless of what happens upstream.</li></ul><p><strong><em>Halfway there: Here’s a shortcut</em></strong></p><p>If Red Flags #1–3 are already showing up in your dependency audit, you don’t have to tackle them manually one by one. Cloudsmith’s <a href="https://docs.cloudsmith.com/repositories/upstreams">upstream proxying and package firewall</a> lets you cache approved versions of upstream packages, enforce origin and vulnerability policies at ingestion, and insulate your builds from upstream changes, all without rearchitecting your existing workflow.</p><p>Worth a look before you get to Red Flags #4 and #5 →</p><h3>Red flag #4: Version ranges that allow automatic upgrades</h3><p><strong>What to look for: </strong>Dependency files using ^, ~, *, or &gt;= version constraints instead of pinned versions. Lockfiles are inconsistently committed or not committed at all.</p><p>This one is so common it barely registers as a risk for most teams until it does.</p><p>Loose version constraints mean your build today may pull a different package version than your build last week. A maintainer can push a breaking or malicious update, and your pipeline will pull it automatically on the next build, with no one on your team making a conscious decision to upgrade.</p><p>Supply chain integrity requires <strong>build reproducibility</strong>. Using pinned package versions ensures you can verify everything you ship, easily recreate historical builds, and provide a meaningful audit trail for compliance and incident response.</p><p><strong>What to do:</strong></p><ul><li>Pin your dependencies to exact versions.</li><li>Commit your lockfiles to version control.</li><li>Use a private artifact repository to cache and proxy approved versions, ensuring your builds remain fully isolated from upstream registry changes, even if the original package is deleted or compromised.</li></ul><h3>Red flag #5: No visibility into transitive dependencies</h3><p><strong>What to look for:</strong> You can list your direct dependencies confidently, but you have no clear picture of what those dependencies depend on, and your security tooling only scans the top level.</p><p>This is the most consequential blind spot in modern software supply chains. When you install a package, you’re not installing one thing; you’re installing a whole dependency tree. A handful of direct dependencies can quietly pull in hundreds of transitive ones.</p><p>When a vulnerability surfaces in a widely used lower-level library, teams with no transitive dependency visibility are left scrambling because they simply didn’t know that library was in their stack. Attackers know this. Dependency confusion attacks and typosquatting are frequently aimed at transitive dependencies precisely because most organizations have zero visibility there.</p><p><strong>What to do:</strong></p><ul><li>Use tooling that resolves and scans your full dependency graph — not just the surface level.</li><li>Generate a Software Bill of Materials (SBOM) to give you a complete, auditable snapshot of everything in your build.</li><li>Pair the SBOM with a private package registry that enforces policy at the point of ingestion.</li><li><a href="https://docs.cloudsmith.com/artifact-management/sbom">Learn how Cloudsmith generates SBOMs automatically →</a></li></ul><h3>The common thread: You can’t secure what you can’t see</h3><p>Every one of these red flags shares the same root: a visibility gap between what your team thinks is in your builds and what’s actually there. Download counts, star ratings, and institutional familiarity are not security controls.</p><p>The teams that stay ahead of supply chain risk are the ones who treat dependency management as a first-class engineering concern, with the same rigor they apply to code review, testing, and deployment. That means:</p><ul><li>Knowing exactly what packages are in your builds, including transitive dependencies.</li><li>Controlling which versions are allowed to enter your pipeline.</li><li>Maintaining an audit trail that captures what was used, when, and where it came from.</li><li>Being able to respond immediately when a package is compromised.</li></ul><p>Treating dependency management as a first-class engineering concern, with the same rigor applied to code review, testing, and deployment, is what separates teams that ship with confidence from teams that discover problems during incidents.</p><h3>How Cloudsmith helps you get there</h3><p>Cloudsmith is a universal artifact management platform built for engineering teams who want full visibility and control over their software supply chain without slowing down how they ship.</p><ul><li><strong>Proxy and cache:</strong> Insulate your builds from upstream deletions and compromises.</li><li><strong>Policy at ingestion: </strong>Block packages by license type, vulnerability severity, or origin before they reach your pipeline.</li><li><strong>Automatic SBOMs: </strong>Maintain a complete, auditable inventory across every build.</li><li><strong>Native integration: </strong>Full support for NPM, PyPI, Maven, Docker, Helm, and 25+ more formats. With Cloudsmith’s multi-format repository support, your entire organisation works from a single, unified artifact store.</li></ul><p>Knowing exactly what’s in your builds, and having the controls to act on that knowledge, is what makes supply chain security a practice, not a checkpoint.</p><p><strong>Ready to get full visibility into your software supply chain? </strong><a href="https://cloudsmith.com/book-a-demo"><strong>Start with Cloudsmith →</strong></a></p><p><em>We publish guides like this regularly for engineering and DevSecOps teams navigating software supply chain security. Follow Cloudsmith to get the next one straight to your feed.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d7435102bb48" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[You have 24 hours to meet CRA. Your software supply chain has no idea]]></title>
            <link>https://cloudsmith.medium.com/you-have-24-hours-to-meet-cra-your-software-supply-chain-has-no-idea-41cf805df135?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/41cf805df135</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cyber-resilience-act]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 11:09:00 GMT</pubDate>
            <atom:updated>2026-04-16T11:09:00.250Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="September 2026: Is Your Software Supply Chain “CRA-Ready”?" src="https://cdn-images-1.medium.com/max/1024/1*rGVIfbmUjcS60nLzsRMmRA.png" /><figcaption>September 2026: Is Your Software Supply Chain “CRA-Ready”?</figcaption></figure><p><strong>September 2026 is no longer on the horizon. It is a live operational reality. And the EU Cyber Resilience Act (CRA) isn’t waiting for your team to figure out the difference between having an SBOM and being compliant.</strong></p><p>There’s a dangerous comfort spreading through engineering teams. It sounds like this:</p><p><em>“We’re fine. We have SBOMs.”</em></p><p>Here is the reality from Cloudsmith’s 2026 Artifact Management Report: <strong>95% of teams generate SBOMs. Only 25% use them to automate gatekeeping.</strong></p><p>That gap between generating a static document and operationalizing the data within it is the <strong>enforcement gap.</strong> It’s where the legal and financial weight of the CRA will hit hardest.</p><h3>The Clock is already running</h3><p>The CRA introduces a hard deadline that manual processes cannot meet. The clock starts when you identify a “severe incident” or an “actively exploited vulnerability”:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/796/1*6GDMOViwjCahZNLbqlMgxA.png" /></figure><p>Twenty-four hours from discovery to a formal, structured report filed with EU cybersecurity authorities. Not to a patch. Not to an investigation. To a <em>report</em>.</p><p><strong><em>“Now ask yourself: how long would that take your team today?”</em></strong></p><p>If the answer involves the words “manually” or “we’d need to loop in multiple teams”, you’re already behind. According to our research, <strong>74% of respondents lack the visibility to produce a comprehensive audit report quickly.</strong> That’s three-quarters of the industry, sleepwalking toward a compliance deadline with a static PDF.</p><h3>Your SBOM is a snapshot. The CRA wants a live feed</h3><p>The CRA doesn’t want a point-in-time PDF. It wants an operational posture. For 75% of teams, an SBOM is a static artifact, a digital receipt stored in a folder. But the moment a new CVE drops or a transitive dependency shifts, that document is stale.</p><p><strong>“CRA-ready teams treat the SBOM as a live data asset.”</strong></p><p>What CRA readiness actually requires is different: an SBOM that is regenerated and verified on every build, continuously reconciled against live vulnerability feeds, and queryable enough that you can answer, “What is affected and where does it live?” in minutes, not after a cross-team investigation. Only 25% of engineering teams currently operate this way.</p><p>Without that operational foundation, you aren’t just failing a security check; you are failing a legal mandate. Visibility and compliance are not the same thing. You can see the iceberg; that doesn’t mean you’ve changed course.</p><h3>The enforcement gap in numbers</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/744/1*XixGBaXgoEuQpEzONc53eA.png" /></figure><h4><strong>Manual compliance doesn’t scale</strong></h4><p>Imagine a vulnerability drops at 2 AM. It affects a transitive dependency four layers deep in one of your packages. By 2 AM the next day, the EU requires a structured report: provenance, version origins, and affected products.</p><p>If your process involves manually cross-referencing package manifests and assembling a report in a spreadsheet, you’re not meeting this window.</p><p>This isn’t hypothetical. <strong>44% of organizations in our survey experienced a real security incident caused by a third-party dependency in the past year.</strong></p><p><em>The vulnerabilities are real. The timeline is real. The gap is real.</em></p><h3>The blueprint for CRA readiness</h3><p>Being CRA-ready means three things:</p><ul><li><strong>Continuous SBOM generation.</strong> Every build produces a verified, queryable SBOM, not a periodic report, but a live data asset reflecting your current artifact state.</li><li><strong>Automated policy enforcement.</strong> When a CVE drops, your system already knows which artifacts are affected and responds, blocking, flagging, and alerting, before a human reads the morning news.</li><li><strong>On-demand audit reporting.</strong> When the 24-hour clock starts, you generate a complete compliance report in minutes. Not hours. Minutes.</li></ul><p><em>The CRA measures whether your systems can act as fast as the threats they face.</em></p><h3>Benchmark your readiness before September</h3><p>The <strong>2026 Artifact Management Report</strong> captures where 505 engineering, security, and platform professionals actually stand on SBOM maturity, supply chain visibility, and CRA readiness.</p><p>Use it to benchmark your team and find your enforcement gap before a regulator does.</p><p><a href="https://cloudsmith.com/campaigns/2026-artifact-management-report"><strong>Download the 2026 Artifact Management Report →</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=41cf805df135" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The dirty secret of AI-generated code: It ships fast and breaks quietly]]></title>
            <link>https://cloudsmith.medium.com/the-dirty-secret-of-ai-generated-code-it-ships-fast-and-breaks-quietly-9b9286499c4e?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b9286499c4e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Thu, 09 Apr 2026 11:49:32 GMT</pubDate>
            <atom:updated>2026-04-09T11:49:32.935Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6OZoNieRN7AqZURfh1p-pg.png" /><figcaption>93% of teams ship AI code. 17% can actually secure it. The rest are gambling</figcaption></figure><p>You adopted AI to move faster. Congratulations, you’re now spending 40 hours a month cleaning up after it.</p><p>That’s not a failure of AI. That’s a failure of the security infrastructure underneath it, one that was built for humans writing code at human speed and is now completely outpaced.</p><h3>The speed you gained came with a bill attached</h3><p>According to <strong>Cloudsmith’s 2026 Artifact Management Report</strong>, based on responses from 505 engineering and security professionals, <strong>93% of organizations</strong> are now using AI to accelerate development. That number skyrocketed from 42% just one year ago. AI-generated code moved from experiment to standard operating procedure almost overnight.</p><p>Here is what hasn’t kept pace: the security infrastructure watching over it.</p><p>Only <strong>17% of organizations</strong> automate scanning and governance specifically for AI-generated code risks. The other 83% rely on security processes designed for a world where humans wrote code at human speed, made human-sized mistakes, and pulled in dependencies one at a time.</p><h3>Meet “slopsquatting”: When hallucinations become attack vectors</h3><p>When an AI coding agent writes code, it doesn’t always reach for existing packages. Sometimes it hallucinates them, fabricating plausible-sounding package names that don’t actually exist in any registry.</p><p>This used to be a minor annoyance, a failed build, a quick manual fix. But attackers figured out the pattern.</p><p><strong>Slopsquatting</strong> is the practice of registering those hallucinated package names before anyone notices. When the AI suggests a fictional package like auth-helper-core and a developer accepts it, the malicious version sitting in the registry is pulled straight into your build.</p><p>The attack surface is no longer just “large”; it’s being actively mapped by adversaries using the same AI tools you are. The average application already depends on over <strong>1,200 libraries</strong>; AI doesn’t make that problem smaller, it accelerates it.</p><h3>The “trust then verify” model is buckling</h3><p>Security teams have long operated on a “trust then verify” posture: ship fast, assume packages are safe, and flag CVEs when they emerge. It works when humans are in the loop, making deliberate choices about what goes into a codebase.</p><p>It does not work when an autonomous agent is doing it at scale, continuously, without stopping to read the room.</p><p>The data shows the strain. <strong>58% of engineering teams now spend 11- 40+ hours per month</strong> just validating and securing AI-generated code. That’s not productivity; that’s a hidden operational tax that eats directly into the speed gains AI was supposed to deliver.</p><p>You didn’t get a 10x developer. You got a developer who now spends a week out of every month cleaning up after one.</p><h3>The enforcement gap nobody wants to talk about</h3><p>Here’s where the report gets uncomfortable.</p><p>Most organizations aren’t actually blind to their vulnerabilities, <strong>79% can identify whether a specific vulnerability affects their environment within six hours of disclosure.</strong> That’s genuinely impressive visibility.</p><p>The problem is what happens next.</p><p>Only <strong>25% of organizations are actually using their security data to automate gatekeeping.</strong> The rest have visibility without enforcement; they can <em>see</em> the threat, they just can’t <em>stop</em> it without significant manual effort.</p><p>When a new CVE drops or a malicious package is flagged, the typical response involves manually assessing exploitability, hunting down downstream dependencies across fragmented toolchains, and defining remediation paths under time pressure. At the human scale, that’s already a stretch. At AI scale, where code volume is surging, and dependency graphs are expanding faster than any team can track, it’s not a process. It’s an aspiration.</p><p><strong>The EU Cyber Resilience Act (CRA) doesn’t care about your aspirations</strong></p><p>Entering into full force in <strong>September 2026</strong>, the CRA is a landmark piece of EU legislation that sets mandatory cybersecurity requirements for all “products with digital elements” sold in the EU. This isn’t just for European companies; if you ship software used in the EU, you are in scope.</p><p>Under the CRA, the “I didn’t know” defense is dead. The act mandates:</p><ul><li><strong>Strict Vulnerability Reporting:</strong> You must notify ENISA (the EU cybersecurity agency) of any exploited vulnerability within <strong>24 hours</strong> of discovery.</li><li><strong>Mandatory Security Support:</strong> Manufacturers must ensure security is “designed-in” and provide automated updates for the product’s lifecycle.</li><li><strong>Heavy Financial Teeth:</strong> Non-compliance can trigger fines up to <strong>€15 million or 2.5% of global turnover</strong>, whichever is higher.</li></ul><p>Only 27% of respondents in our report feel “very confident” they could pass an unexpected supply chain audit today. The other 73% are looking at a collision course between AI-generated code volume and a legal mandate that demands near-instant transparency. When the CRA hits, “cleaning up after the AI” won’t just be an operational tax; it will be a legal liability.</p><h3>AI models are artifacts, too. Most teams haven’t caught up to that.</h3><p>There is a second AI problem hiding inside the first: <strong>Governance fragmentation.</strong></p><p>As AI models themselves become a core part of the software supply chain, organizations are storing and governing them very differently from other artifact types. The most commonly used open-source models come from Google (Gemma, adopted by 69%) and Meta (LLaMA variants, at 47%). The familiarity of the source creates a false sense of security, which the report calls <strong>“reputational security.”</strong></p><p>Because the model comes from a trusted vendor, it doesn’t go through the same verification pipeline as a Docker container or a Python package. Checksums get verified. Internal safety? Often not inspected at all.</p><p>The result: <strong>only 12% of organizations manage AI models using the same security policies and provenance tracking they apply to traditional binaries.</strong> The artifact type that is arguably the most opaque and consequential is the one that is governed the least rigorously.</p><p>Isolating AI models from standard artifact management doesn’t make them safer. It just makes them invisible to the controls you’ve already built.</p><h3>2026 Strategy: Move from “trust then verify” to “verify then trust”.</h3><p>The “trust then verify” model made sense when the bottleneck was development speed and the threat landscape moved at a human pace. Neither of those things is true anymore.</p><p>The shift organizations need to make is to a <strong>“verify then trust”</strong> posture, one where governance policies run <em>before</em> code reaches the development environment, not after it’s already in production. Where AI-generated dependencies are scanned at ingestion. Where AI models are treated as first-class artifacts with the same provenance requirements as everything else in the pipeline.</p><p>Critically, this doesn’t have to mean slowing down. The whole point of shifting enforcement left and automating it is that developers keep moving fast; they just move fast inside guardrails they don’t have to think about.</p><p>The teams winning at this aren’t spending more time on security. They’ve consolidated their toolchain into a unified control plane that applies consistent policies across every artifact type, code, containers, packages, and models, so that automation handles what used to require a human.</p><h3>What to do with this</h3><p>If you’re one of the 83% still relying on manual processes to govern AI-generated code, the window to get ahead of this is narrowing. Supply chain attacks are not slowing down. The CRA is not moving its timeline. And AI adoption inside your org almost certainly isn’t reversing.</p><p>The 2026 Artifact Management Report maps exactly where the enforcement gaps are, what the highest-risk failure modes look like, and what the organizations ahead of this are doing differently.</p><p><strong>→ </strong><a href="https://cloudsmith.com/campaigns/2026-artifact-management-report"><strong>Download the 2026 Artifact Management Report</strong></a></p><p>It’s the data your security roadmap needs. And it’s a significantly better use of your time than the 40 hours a month you might currently be spending cleaning up what your AI shipped without asking.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b9286499c4e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I asked AI to code the backend. It suggested a malicious package that didn’t exist (Yet)]]></title>
            <link>https://cloudsmith.medium.com/i-asked-ai-to-code-the-backend-it-suggested-a-malicious-package-that-didnt-exist-yet-07b43a9e27c8?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/07b43a9e27c8</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[supply-chain]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Wed, 01 Apr 2026 11:04:48 GMT</pubDate>
            <atom:updated>2026-04-01T11:04:48.688Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="AI Hallucinations in Software Supply Chain: How to Prevent Attacks" src="https://cdn-images-1.medium.com/max/1024/1*pZ28pWH7_1OfawTLh48uVw.png" /></figure><p>“I didn’t install anything malicious. The AI suggested it.”</p><p>That might soon become the most expensive sentence in your next post-mortem.</p><p>Last week, while spinning up a quick microservice, I asked my AI assistant for a standard library to handle a specific type of JWT validation in a niche Node.js framework. The code it spat out looked perfect. The logic was sound. But there was a line in the generated <em>README</em> instructions:</p><p>&lt;npm install express-jwt-auth-validator&gt;</p><p>I ran the command: <strong>404 error.</strong> The package didn’t exist.</p><p>I laughed it off as a classic AI hallucination. But then the realization hit me: if I’m seeing this hallucination, then thousands of other developers are, too. And in the world of cybersecurity, a “statistically likely” hallucination is just a pre-made target for an attacker.</p><p>Welcome to the era of <strong>slopsquatting.</strong></p><h3><strong>The AI coding boom’s dark matter</strong></h3><p>AI coding tools such as ChatGPT, Copilot, and Claude have become our force multipliers. We use them to ship faster, but we’re ignoring a fundamental truth: <strong>AI doesn’t verify dependencies; it predicts them.</strong></p><p>When an LLM suggests a library, it isn’t searching a live database of the npm or PyPI registries. It predicts the next most likely string of text using billions of parameters. If a name sounds right, like secure-data-parser or react-native-auth-flow, the AI will suggest it with 100% confidence because it <em>should</em> exist.</p><p>This creates a<strong> “phantom” </strong>dependency, a hole in the software supply chain that is currently empty but perfectly shaped for a malicious payload.</p><h3>How a hallucination becomes a breach</h3><p>This is how a phantom package becomes a weaponized exploit:</p><ol><li><strong>The hallucination:</strong> An AI consistently suggests a non-existent package to developers globally.</li><li><strong>The observation:</strong> Attackers monitor LLM outputs or common developer search trends for these specific phantom names.</li><li><strong>The registration:</strong> An attacker registers one of these phantom packages, like express-jwt-auth-validator, on a public registry.</li><li><strong>The payload:</strong> They push a version that looks functional but contains a hidden postinstall script to exfiltrate your .env files or inject a backdoor.</li><li><strong>The execution:</strong> The next developer who asks the AI for help installs a “hallucinated” package that is now very real and very malicious.</li></ol><p>Traditional security tools, such as Software Composition Analysis (SCA), scan for <em>known</em> vulnerabilities in <em>existing</em> packages. They aren’t designed to stop a package that was literally predicted into existence by an AI five minutes ago.</p><h3>Why Cloudsmith is the firewall for AI hallucinations</h3><p>You don’t have to ban AI to stay safe. You just have to stop letting your developers pull AI-suggested packages directly from the public internet. To survive the era of AI-<strong>powered</strong> development, your infrastructure needs to move from “<strong>public-first</strong>” to <strong>“private-first”</strong>.</p><p>This is where <strong>Cloudsmith</strong> comes in. It acts as the immune system for your software supply chain.</p><h3>1. The walled garden strategy</h3><p>Pulling directly from public registries is now a high-risk activity. Cloudsmith lets you create a <strong>private artifact repository</strong> that serves as a secure buffer. Your CI/CD and your developers pull from your curated Cloudsmith repo. If an AI suggests a hallucinated package, it won’t be in your registry. The build fails. The breach is stopped before it starts.</p><h3>2. Automated policy enforcement</h3><p>Security shouldn’t rely on a developer’s memory. With <strong>Cloudsmith’s enterprise policy manager</strong>, you can set guardrails that an AI can’t bypass. Every package that enters your ecosystem is automatically scanned for malware and subjected to policy enforcement before it ever reaches your developers. On top of that, you can create policies to automatically block packages that:</p><ul><li><strong>Were registered in the last 24 hours:</strong> A prime window for slopsquatting. (Or any other cooldown period that protects your systems and doesn’t slow developers down.)</li><li><strong>Lack a verified maintainer:</strong> Ensuring the human element remains in the loop.</li><li><strong>Fail signature verification:</strong> Ensuring the code is exactly what it claims to be.</li></ul><h3>3. Full supply chain visibility</h3><p>In 2026, <strong>agentic workflows,</strong> where AI agents write and deploy code autonomously, are becoming the norm. You cannot secure what you cannot see. Cloudsmith provides a single control plane to see every dependency entering your ecosystem, whether it was added by a human or a hallucinating bot.</p><h3>The shift: Trust is the new infrastructure</h3><p>For a decade, the industry focused on writing secure code. Today, the battle has shifted to <strong>secure provenance.</strong></p><p>Trusting an AI with your dependencies without a management layer like <strong>Cloudsmith</strong> is like letting a stranger grocery shop for your family; eventually, something toxic could end up in the cart.</p><p>In the age of AI, “Move Fast and Break Things” has become “Move Fast and Hallucinate Breaches”. It’s time to put a filter between the AI’s imagination and your production environment.</p><h3>Final thoughts</h3><p>The scariest thing about AI-generated dependencies? They don’t feel like a threat. They feel like a shortcut.</p><p>The next time your AI tool suggests a perfect new library, don’t just copy-paste. Verify the source. Better yet, make sure your infrastructure is powered by a platform that verifies it for you.</p><p><strong>Is your supply chain protected from AI hallucinations?</strong><a href="https://cloudsmith.com/"><strong> </strong></a><a href="https://cloudsmith.com/blog/slopsquatting-and-typosquatting-how-to-detect-ai-hallucinated-malicious-packages"><strong>Check out how Cloudsmith secures your artifacts here.</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=07b43a9e27c8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Is AI quietly making your software supply chain less secure?]]></title>
            <link>https://cloudsmith.medium.com/is-ai-quietly-making-your-software-supply-chain-less-secure-e1364de33f9a?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/e1364de33f9a</guid>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[supply-chain]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Tue, 24 Mar 2026 15:45:22 GMT</pubDate>
            <atom:updated>2026-03-24T15:46:39.445Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>Is AI Quietly Making Your Software Supply Chain Less Secure?</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QxzBMnPE0KPxqXido8nqCw.png" /><figcaption>The rise of “slopsquatting”</figcaption></figure><p>There is a “ghost in the machine” starting to appear in our codebases.</p><p>It looks like a perfectly formatted pull request. The code is clean, follows every stylistic rule in the book, and looks like it was written by your best senior engineer. It’s elegant, efficient, and was generated in roughly three seconds. But underneath the surface, it’s behaving like a Trojan horse for the modern software supply chain.</p><p>Working at <strong>Cloudsmith</strong>, we have a front-row seat to the billions of artifacts and dependencies moving through the global software ecosystem. Lately, we’ve noticed a quiet, unsettling shift. As vibe coding becomes the norm, where developers ship code because the AI output <em>feels</em> right, we are accidentally trading long-term auditability for short-term velocity.</p><p>The truth is, AI isn’t just writing your code; it’s quietly rewriting your risk profile.</p><h3>The rise of “slopsquatting”</h3><p>The biggest threat to your supply chain right now isn’t a sophisticated nation-state actor; it’s a <strong>hallucination</strong>.</p><p>Security researchers recently identified a trend called <strong>slopsquatting</strong>. When an LLM generates code, it occasionally suggests a library or a package that doesn’t actually exist. It sounds plausible, something like fast-auth-utils or react-dom-validator.</p><p>The danger? Hackers are now monitoring common AI hallucination patterns. When an AI frequently suggests a fake package name, attackers register that exact name in public registries like npm or PyPI and fill it with malicious payloads.</p><p>The developer trusts the AI, the AI trusts the hallucination, and your CI/CD pipeline pulls the poison directly into your production environment.</p><h3>45% of AI code is born vulnerable</h3><p>It’s easy to blame the AI, but the model is just a mirror. It was trained on decades of legacy code, much of which was written before modern security standards existed.</p><p>According to recent 2025–2026 industry reports, nearly <strong>45% of AI-generated code</strong> contains at least one high-severity vulnerability. The most common offenders?</p><ul><li><strong>SQL injection:</strong> AI loves the “shortest path” to a solution, which often means skipping input sanitization.</li><li><strong>Insecure defaults:</strong> Hardcoded keys or permissive CORS policies that look “ready for testing” but end up in production.</li><li><strong>Shadow dependencies:</strong> AI pulling in outdated, unmaintained libraries that haven’t seen a patch since 2022.</li></ul><h3>The “oversight gap”</h3><p>The real risk isn’t the AI, it’s the <strong>human.</strong> As we rely more on automated assistants, our “manual review muscles” are atrophying. When we write a line of code, we (usually) understand why it’s there. When an AI writes 50 lines, the human reviewer often skims for logic but misses the subtle security debt hidden in the middle.</p><p>We are trading <strong>auditability</strong> for <strong>velocity</strong>. And in a world of regulated software, that is a dangerous trade to make.</p><h3>How to stay fast (without getting hacked)</h3><p>We aren’t going to stop using AI. The productivity gains are too high. But we do need to change how we treat the artifacts it produces.</p><ol><li><strong>Trust, but verify (with policy):</strong> You shouldn’t allow any package, AI-suggested or otherwise, into your build unless it has been vetted against a centralized, internal library of approved and verified dependencies.</li><li><strong>AIBOMs are the new SBOMs:</strong> You already use a Software Bill of Materials. In 2026, you need an <strong>AI-BOM</strong> to track which parts of your codebase were machine-generated so you can apply stricter auditing to those modules.</li><li><strong>The “package firewall”:</strong> Your artifact repository shouldn’t just be an artifact graveyard or a dumping ground for unverified code; it should be a shield. If a hallucinated or slopsquatted package tries to enter your ecosystem, your repository should be the one to say “No.”</li></ol><h3>The real cost of vibe coding</h3><p>AI is moving faster than our security policies can keep up. We are currently in the honeymoon phase of machine-accelerated development, where the productivity gains feel like magic.</p><p>But as we’ve seen at <strong>Cloudsmith</strong>, the magic comes with a footprint. The artifacts being called into your environment today aren’t just lines of code; they are the new frontier of the software supply chain. Whether it’s a hallucinated package or an insecure default, the ghost in the machine is real, and it’s already here.</p><p>The industry-wide challenge for 2026 isn’t <strong>“How do we code faster?”</strong> but <strong>“How do we verify at the speed of thought?”</strong></p><p>The software supply chain was never meant to be a black box. If we lose the ability to audit what we ship, we lose the ability to trust our own products. The question isn’t whether you’ll use AI, it’s whether you’ll be the one who actually owns the results.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e1364de33f9a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[13 KubeCon Europe 2026 sessions not to miss]]></title>
            <link>https://cloudsmith.medium.com/13-kubecon-europe-2026-sessions-not-to-miss-701a79a4c930?source=rss-2f200e7ca9d3------2</link>
            <guid isPermaLink="false">https://medium.com/p/701a79a4c930</guid>
            <category><![CDATA[cloud-native]]></category>
            <category><![CDATA[platform-engineering]]></category>
            <category><![CDATA[kubecon2026]]></category>
            <category><![CDATA[kubeconeu]]></category>
            <category><![CDATA[kubecon]]></category>
            <dc:creator><![CDATA[Cloudsmith]]></dc:creator>
            <pubDate>Tue, 24 Feb 2026 10:51:37 GMT</pubDate>
            <atom:updated>2026-02-24T10:51:37.467Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GwZQPZN1WlSb6ua1T8wyKg.png" /><figcaption>KubeCon Europe 2026</figcaption></figure><p>Thousands of platform engineers and security experts are about to gather for KubeCon + CloudNativeCon Europe 2026, making it the year’s definitive cloud-native event. To help you cut through the noise of a packed agenda, we’ve identified the standout sessions you can’t afford to miss. Whether you’re looking for high-level strategy or interactive workshops, here is our curated roadmap to ensure you get the most out of your time on the ground.</p><p><strong>📍 RAI Amsterdam (Europaplein 24, 1078 GZ Amsterdam, Netherlands)<br>🗓️ March 24–26, 2026</strong></p><h3>1. SBOOM: Making SBOMs Play Together</h3><ul><li><strong>Date/Time:</strong> March 26, 14:30–15:00 CET</li><li><strong>Location:</strong> Hall 8 | Room F</li><li><strong>Theme:</strong> Security</li></ul><p>In this session, research engineers <strong>Jacopo Bufalino</strong> (CNAM) and <strong>Agathe Blaise</strong> (Thales SIX GTS France) will dive into the messy reality of the Software Bill of Materials (SBOM). They plan to pull back the curtain on why current open-source and cloud-based tools (despite their promises) frequently produce conflicting package lists and inconsistent vulnerability reports when scanning complex container images.</p><p>This talk is particularly electrifying for the security community because it tackles the compliance anxiety triggered by the EU’s Cyber Resilience Act (CRA). As the CRA shifts SBOMs from a nice-to-have transparency initiative to a legal necessity for software lifecycles, the industry is waking up to a major problem: inconsistent data. By dissecting the technical roots of these discrepancies, Bufalino and Blaise aren’t just pointing out flaws; they will also be providing a roadmap for developers to ensure their tooling is actually CRA-compliant. For anyone navigating the shift toward a transparent and trustworthy software supply chain, this session offers the technical clarity needed to turn these metaphorical bombs into secure builds.</p><p><a href="https://kccnceu2026.sched.com/event/2CW7Q/sbm-making-sboms-play-together-jacopo-bufalino-cnam-agathe-blaise-thales-six-gts-france"><strong>View session details</strong></a></p><h3>2. Modelpack: Standardising the packaging and distribution of AI/ML Models</h3><ul><li><strong>Date/Time:</strong> March 23, 15:10–15:15 CET</li><li><strong>Location:</strong> Elicium 2</li><li><strong>Theme:</strong> AI + ML</li></ul><p>In this high-impact lightning talk, <strong>Andrew Block</strong>, a Distinguished Architect at Red Hat, tackles one of the most pressing bottlenecks in modern infrastructure: the chaotic fragmentation of AI/ML integration within the cloud-native ecosystem. Block, a seasoned expert in helping organisations scale open-source solutions, will break down how the current Wild West of competing formats and runtimes actively stifles innovation. He will introduce <a href="https://modelpack.org/"><strong>ModelPack</strong></a>, a CNCF Sandbox project designed to act as the universal translator for AI/ML artifacts, allowing them to finally speak the same language as established tools like Kubernetes, ORAS, and Harbor.</p><p>The security and DevOps communities are closely watching this session because it addresses the Day 2 operations nightmare of managing AI at scale. As AI moves from experimental notebooks to production-grade Kubernetes clusters, the lack of standardisation creates massive technical debt and security blind spots. Block’s exploration of ModelPack and emerging artifact standards offers a glimpse into a future where AI models are managed with the same consistency and rigour as container images. For anyone struggling to bridge the gap between data science and cloud-native engineering, this talk provides the blueprint for a unified, scalable, and manageable AI supply chain.</p><p><a href="https://kccnceu2026.sched.com/event/2EFyi/project-lightning-talk-modelpack-standardizing-the-packaging-and-distribution-of-aiml-models-andrew-block-maintainer"><strong>View session details</strong></a></p><h3>3. Kubernetes security at Shopify scale: Automating security across an infrastructure monrepo</h3><ul><li><strong>Date/Time:</strong> March 25, 16:00–16:30 CET</li><li><strong>Location:</strong> Hall 8 | Room F</li><li><strong>Theme:</strong> Security</li></ul><p>In this session, Senior Infrastructure Security Engineers <strong>Jie Wu</strong> and <strong>Pulkit Garg</strong> will pull back the curtain on how one of the world’s largest e-commerce platforms secures its massive infrastructure monorepo. Drawing on their deep backgrounds in cloud defense and 5G network security, Wu and Garg will detail the high-stakes challenge of managing thousands of services, where a single misconfiguration could affect millions of merchants. They will demonstrate how Shopify moved beyond manual checkbox security by building an automated pipeline that combines Static Analysis with Semgrep and the Open Policy Agent (OPA) for real-time policy enforcement.</p><p>This talk is a must-attend for the security community because it provides a battle-tested blueprint for resolving the velocity-versus-security dilemma we’re all too familiar with. While many orgs struggle with friction between DevOps and Security teams, Shopify’s approach shows how to bake guardrails directly into the developer workflow without slowing down deployment. Attendees will gain rare insights into the so-called rough patches of scaling open-source security tools and leave with a practical framework for automating risk detection across complex, high-traffic Kubernetes environments.</p><p><a href="https://kccnceu2026.sched.com/event/2CW3b/kubernetes-security-at-shopify-scale-automating-security-across-an-infrastructure-monorepo-jie-wu-pulkit-garg-shopify"><strong>View session details</strong></a></p><h3>4. To upstream or not? Why becoming the maintainer of your dependencies matters</h3><ul><li><strong>Date/Time:</strong> March 26, 11:00–11:30 CET</li><li><strong>Location:</strong> Hall 7 | Room B</li><li><strong>Theme:</strong> Cloud Native Experience</li></ul><p>In the session, <strong>Christos Markou</strong>, a Principal Software Engineer at Elastic and an OpenTelemetry code owner, dives into the strategic heart of modern software engineering. Markou moves beyond the typical <em>“open source is good for the soul”</em> narrative to present a pragmatic, business-first case for active maintenance. By sharing a high-stakes story of an OTel component saved from the brink of deprecation, he demonstrates how moving from a passive consumer to an active contributor transformed a potential technical debt nightmare into a rapid-response tool for solving critical customer issues.</p><p>This talk is a vital wake-up call for the security and observability communities because it addresses the growing crisis of software supply chain sustainability. In an era where a single abandoned dependency can lead to massive security vulnerabilities or operational outages, Markou provides a concrete example to share with your manager on why upstreaming isn’t just altruism; it also can be treated as sensible risk management. For teams building on CNCF projects, this session offers a masterclass in how staying present in the ecosystem pays dividends in architectural control, security posture, and long-term engineering velocity.</p><p><a href="https://kccnceu2026.sched.com/event/2CW5U/to-upstream-or-not-why-becoming-the-maintainer-of-your-dependencies-matters-christos-markou-elastic"><strong>View session details</strong></a></p><h3>5. How to add a new language feature to OPA</h3><ul><li><strong>Date/Time:</strong> March 23, 10:20–10:25 CET</li><li><strong>Location:</strong> Elicium 2</li><li><strong>Theme:</strong> Open Policy Agent</li></ul><p>Adding features to a widely adopted language like Rego involves more than just a simple pull request. In this lightning session, Apple’s <strong>Charlie Egan</strong> explores the full-stack impact of introducing new syntax, from the initial parser modifications to updating the broader ecosystem of editor tools. Learn the specific engineering hurdles involved in upgrading a mature security project without disrupting the developer experience. At Cloudsmith, we’re huge fans of Open Policy Agent and Rego, so we can’t wait to hear what Charlie has in store for us.</p><p><a href="https://kccnceu2026.sched.com/event/2EFwy/project-lightning-talk-how-to-add-a-new-language-feature-to-opa-charlie-egan-maintainer"><strong>View session details</strong></a></p><h3>6. Platform Engineering 2.0: Just-enough Kubernetes and AI-native DevOps</h3><ul><li><strong>Date/Time:</strong> March 26, 15:15–15:45 CET</li><li><strong>Location:</strong> Hall 11</li><li><strong>Theme:</strong> Platform Engineering</li></ul><p>In the session, <strong>Shweta Vohra</strong>, a Lead Architect at <strong>Booking.com</strong> and a seasoned authority on platform patterns, tackles the industry’s growing complexity bloat. Drawing from 20 years of experience and the literal scars of re-architecting internal platforms at one of the world’s largest travel sites, Vohra argues that the future of infrastructure isn’t about building more, it’s really about building just enough. She will break down how to transition from heavy, static systems to lean, right-sized architectures using k3s, Gateway API, and ambient mesh, while layering in AIOps via Kubeflow and Prometheus to transform raw automation into true system intelligence.</p><p>This talk is a critical touchstone for the DevOps and platform communities because it offers a pragmatic exit ramp from the scale-at-all-costs mentality that leads to massive cloud waste and operational blindness. Vohra’s AI-native approach is a blueprint for creating self-optimising ecosystems that align infrastructure directly with business value. For engineering leaders and architects feeling the weight of over-engineered clusters, this session aims to be a masterclass in staying lean, staying smart, and evolving platforms to be adaptive rather than just additive.</p><p><a href="https://kccnceu2026.sched.com/event/2CW80/platform-engineering-20-just-enough-kubernetes-and-ai-native-devops-shweta-vohra-bookingcom"><strong>View session details</strong></a></p><h3>7. AI Agents &amp; Platform Engineering: Efficiency boost or new source of trouble?</h3><ul><li><strong>Date/Time:</strong> March 26, 11:45–12:15 CET</li><li><strong>Location:</strong> Auditorium</li><li><strong>Theme:</strong> AI + ML</li></ul><p>In the blockbuster panel, an elite group of industry leaders from <strong>Cisco</strong>, <strong>Red Hat</strong>, <strong>AWS</strong>, <strong>Solo.io</strong>, and the <strong>United Nations</strong> converges to debate the most disruptive shift in DevOps history. Featuring heavyweights like from the industry, this vendor-neutral discussion moves past the hype to ask the hard questions: Can non-deterministic AI agents actually keep pace with AI-accelerated developers, or are we just inviting unpredictable trouble into our production clusters? The panel will tackle the minimum viable platform foundation required for AI success, the shifting cost implications of agents in production, and how to build human trust in systems that don’t always behave the same way twice.</p><p>This session is arguably the must-see event for the modern platform engineer because it addresses the looming intelligence gap in infrastructure. As developers pump out code at record speeds, the platform team is traditionally treated as the bottleneck. This panel explores whether AI agents are the ultimate scaling solution or a new category of technical debt. By bringing together perspectives from global tech giants and international organisations, attendees will walk away with a high-level framework for defining golden metrics for AI effectiveness and a roadmap for navigating the transition from static automation to dynamic, agent-led collaboration.</p><p><a href="https://kccnceu2026.sched.com/event/2CW6V/ai-agents-platform-engineering-efficiency-boost-or-new-source-of-trouble-hasith-kalpage-cisco-vincent-caldeira-red-hat-sara-qasmi-united-nations-idit-levine-soloio-carlos-santana-aws"><strong>View session details</strong></a></p><h3>8. In Falco’s nest: The evolution of cloud native runtime security</h3><ul><li><strong>Date/Time:</strong> March 24, 12:00–12:30 CET</li><li><strong>Location:</strong> G102–103</li><li><strong>Theme:</strong> Security</li></ul><p>In the maintainer track session, Falco contributors <strong>Iacopo Rozzo</strong> (Sysdig) and <strong>Aldo Lacuku</strong> (Kong) provide a deep-dive update on the CNCF’s de facto standard for runtime threat detection. The core of their presentation focuses on the highly anticipated Falco Operator, a game-changer designed to automate the deployment and management of security sensors across massive, distributed clusters, effectively lowering the barrier to entry for enterprise-scale security.</p><p>For the security community, this talk is a vital look at the roadmap for high-throughput runtime defense. As cloud-native environments become more complex and data-heavy, traditional security tools often struggle with performance overhead; Rozzo and Lacuku will demonstrate the specific optimisations that allow Falco to maintain deep visibility without sacrificing system reliability. Beyond the code, this session also is a great opportunity to see how the Falco ecosystem is integrating with the broader CNCF landscape, offering attendees a first look at the features that will define Cloud Detection &amp; Response (CDR) in the coming year.</p><p><a href="https://kccnceu2026.sched.com/event/2EF6W/in-falcos-nest-the-evolution-of-cloud-native-runtime-security-iacopo-rozzo-sysdig-aldo-lacuku-kong-inc"><strong>View session details</strong></a></p><h3>9. What LLMs do, and don’t, know about securing Kubernetes</h3><ul><li><strong>Date/Time:</strong> March 24, 14:30–15:00 CET</li><li><strong>Location:</strong> Hall 11</li><li><strong>Theme:</strong> Security</li></ul><p>In the session, <strong>Rory McCune</strong>, a Senior Security Researcher at <strong>Datadog</strong>, explores the practical (and often perilous) intersection of Generative AI and cluster orchestration. McCune, a veteran KubeCon speaker and a foundational voice in container security, will present the results of his deep-dive research into whether Large Language Models (LLMs) can actually be trusted with sensitive credentials. Rather than just discussing theory, Rory will aim to demonstrate how LLMs handle specific Kubernetes security tasks, revealing where they provide genuine architectural insight and where they tend to hallucinate dangerous misconfigurations that could leave an organisation exposed.</p><p>This talk is a critical reality check for the security community as AI-assisted DevOps moves from a trend to a standard operating procedure. McCune will break down how advanced techniques like improved prompting and <a href="https://www.promptingguide.ai/techniques/cot"><strong>chain-of-thought</strong></a> reasoning can significantly shift the safety of an LLM’s output, while also highlighting the so-called no-go zones where human expertise remains non-negotiable. For security architects and engineers, this session provides a vital framework for auditing AI-generated IaC templates, ensuring that the speed of AI doesn’t come at the cost of a catastrophic security breach.</p><p><a href="https://kccnceu2026.sched.com/event/2CVyo/what-llms-do-and-dont-know-about-securing-kubernetes-rory-mccune-datadog"><strong>View session details</strong></a></p><h3>10. Audit-ready Kubernetes: How Chase UK leveraged policy-as-code for continuous compliance</h3><ul><li><strong>Date/Time:</strong> March 25, 14:15–14:45 CET</li><li><strong>Location:</strong> Hall 7 | Room C</li><li><strong>Theme:</strong> Security</li></ul><p>In the session, <strong>Nischay Goyal</strong>, VP of Cloud Platform Engineering at <strong>JPMorgan Chase</strong>, and <strong>Jim Bugwadia</strong>, CEO of <strong>Nirmata</strong> and co-creator of Kyverno, provide a rare under-the-hood look at building a cloud platform in the high-stakes world of regulated retail banking. Goyal, who manages the end-to-end Kubernetes ecosystem for Chase UK, will detail how his team transformed a massive compliance undertaking into a streamlined, automated engine. By leveraging Kyverno, OpenReports, and Grafana, they successfully shifted security left, enabling their backend engineers to ship code at speed while maintaining a real-time policy-as-code safety net that meets stringent financial regulatory requirements.</p><p>This talk is a beacon for the security and compliance communities because it addresses the ultimate white whale of enterprise DevOps: reducing audit times from weeks to minutes. Bugwadia’s deep perspective as a Kubernetes <a href="https://kubernetes.io/blog/2025/10/18/wg-policy-spotlight-2025/"><strong>Policy Working Group</strong></a> co-chair, combined with Goyal’s real-world production experience at JPMorgan Chase, offers a definitive guide on how to empower security teams to write independent policies without bottlenecking development. For any platform team operating in a regulated sector (or simply running mission-critical workloads) this session provides a proven framework for turning compliance from a manual, fear-driven process into a better, more transparent, and scalable process.</p><p><a href="https://kccnceu2026.sched.com/event/2CW2L/audit-ready-kubernetes-how-chase-uk-leveraged-policy-as-code-for-continuous-compliance-jim-bugwadia-nirmata-nischay-goyal-jp-morgan-chase"><strong>View session details</strong></a></p><h3>11. Sandbox Operator: Enabling session-aware, efficient MCP tool execution in Kubernetes</h3><ul><li><strong>Date/Time:</strong> March 25, 11:00–11:30 CET</li><li><strong>Location:</strong> Auditorium</li><li><strong>Theme:</strong> AI + ML</li></ul><p><strong>Alibaba</strong> Cloud engineers <strong>Mingshan Zhao</strong> and <strong>Zhen Zhang</strong> tackle the next massive infrastructure hurdle for AI: the Model Context Protocol (MCP). As maintainers of the OpenKruise project with experience managing Alibaba’s million-container scheduling system, Zhao and Zhang are uniquely qualified to address the so-called “session explosion” problem. They will introduce the Sandbox Operator, a specialised Kubernetes controller designed to manage the lifecycle of AI tool executions without the massive resource waste or state loss typical of traditional Pod deployments.</p><p>This talk is a game-changer for the cloud-native community because it solves the sparse invocation dilemma, where these AI tools often sit idle, eating up expensive cluster resources while waiting for a user’s next prompt. By integrating cutting-edge Checkpoint/Snapshot mechanisms, the Sandbox Operator allows tools to be paused and resumed without losing the conversation&#39;s state. For platform engineers and AI architects, this session shows how developers can efficiently scale AI agents to hundreds of thousands of concurrent users while keeping infrastructure costs low and user context intact.</p><p><a href="https://kccnceu2026.sched.com/event/2CW1c/sandbox-operator-enabling-session-aware-efficient-mcp-tool-execution-in-kubernetes-mingshan-zhao-zhen-zhang-alibaba"><strong>View session details</strong></a></p><h3>12. Saxo service blueprint: Bridging legacy and modern world with Kubernetes operators</h3><ul><li><strong>Date/Time:</strong> March 25, 16:00–16:30 CET</li><li><strong>Location:</strong> Hall 11</li><li><strong>Theme:</strong> Platform Engineering</li></ul><p>In the session, <strong>Jinhong Brejnholt</strong>, Chief Cloud Architect at <strong>Saxo Bank</strong>, presents a sophisticated solution to the two-speed IT problem facing large enterprises. As a certified Kubestronaut and a leader in the Danish cloud-native community, Brejnholt details how Saxo Bank moved beyond the bottleneck of manual ticketing for DNS, certificates, and load balancing. She will showcase the Saxo Service Blueprint, a platform powered by Kubernetes Operators that exports the speed and reliability of GitOps to traditional VM-based applications, effectively unifying legacy infrastructure with modern cloud-native workflows.</p><p>This talk is particularly compelling for the platform engineering community because it addresses the last mile of digital transformation: managing the heavy, non-containerised dependencies that still power most financial institutions. Brejnholt will share how this architecture has saved thousands of developer hours and significantly bolstered disaster recovery capabilities for both cloud-native and legacy stacks. For architects struggling to reconcile modern Kubernetes automation with entrenched enterprise systems, this session offers some valuable real-world experiences and a roadmap for other banks to accelerate delivery across the entire organisational footprint.</p><p><a href="https://kccnceu2026.sched.com/event/2CW4E/saxo-service-blueprint-bridging-legacy-and-modern-world-with-kubernetes-operators-jinhong-brejnholt-saxo-bank"><strong>View session details</strong></a></p><h3>13. Optimising error recovery for cost-efficient distributed AI model training with Kubernetes</h3><ul><li><strong>Date/Time:</strong> March 26, 14:30–15:00 CET</li><li><strong>Location:</strong> Elicium 2</li><li><strong>Theme:</strong> AI + ML</li></ul><p>For our final recommended talk, the academic precision of <strong>Radostin Stoyanov</strong> (undergoing his PhD at the <strong>University of Oxford</strong>) meets the enterprise-scale expertise of <strong>Andrey Velichkevich</strong> (AI Engineer at <strong>Apple</strong> &amp; Kubeflow Steering Committee member). Together, they aim to address the GPU tax that plagues modern AI development: the massive costs incurred when long-running training jobs fail or when idle interactive workloads, like Jupyter notebooks, waste expensive compute resources. The duo will present a breakthrough approach that uses transparent GPU checkpointing, integrated with Kubernetes-native APIs such as Kueue, JobSet, and TrainJob, to automatically capture and restore the state of training jobs, ensuring no progress is lost during failures or preemptions.</p><p>This talk is really exciting for the ML and Platform communities because it unlocks the holy grail of AI infrastructure: the ability to run reliable, high-stakes model training on preemptible spot instances. By moving checkpointing from the application layer to the infrastructure layer, Stoyanov and Velichkevich show how organisations can slash their cloud bills by up to 90% without risking their training timelines. For anyone tasked with scaling AI workloads while maintaining strict cost-efficiency and cluster utilisation, this session offers a sophisticated technical roadmap for turning unstable, expensive hardware into a resilient, high-performance training engine.</p><p><a href="https://kccnceu2026.sched.com/event/2CW7Z/optimizing-error-recovery-for-cost-efficient-distributed-ai-model-training-with-kubernetes-radostin-stoyanov-university-of-oxford-andrey-velichkevich-apple"><strong>View session details</strong></a></p><h4>Secure software starts at Booth 570</h4><p>Cloudsmith will be at booth 570, and we are bringing plenty of action to keep you busy. If you are ready to put your security skills to the test, join our capture-the-flag workshop: Stop the AGI apocalypse. You will hunt hidden malware, lock down LLM weights, and earn a shot at some great prizes. We also have a lineup of fun events to help you unwind, from canal cruises and kickoff drinks to beach-themed parties. <br>Check out the full schedule of Cloudsmith activities here: <a href="https://cloudsmith.com/events/in-person-events/kceu26"><strong>https://cloudsmith.com/events/in-person-events/kceu26</strong></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=701a79a4c930" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>