AI News: Fifty Dollars a Break-In

Anthropic filed its S-1 with the SEC on Monday morning, applying to trade on the Nasdaq Global Select Market under the ticker ANT, with 16.29 billion dollars of first-half revenue, an 80.7 billion dollar net loss that is mostly an accounting artifact, and more than 450 billion dollars of compute obligations behind it.

PHP maintainers shipped emergency patches for two flaws that let an attacker take over a server with no login, and security firm Calif found the first by pointing OpenAI’s GPT-5.6 Sol at a codebase that had carried the bug for about a decade. NorForge published an account of a single operator whose AI agents breached more than 100 companies, Cummins and Deere and SharkNinja among them, at roughly 50 dollars a target.

Axios reported that JPMorgan, Bank of America, Morgan Stanley, and Citi are restricting or blocking third-party agents found operating inside customer accounts, with one executive calling the situation horrifying.

Anthropic set aside 5 percent of its equity for a Public Benefit Fund worth about 48 billion dollars, then published data showing junior roles falling from 27 percent of its own technical hires to 9 percent.

Sam Altman gave OpenAI staff three possible endings for the safety race and a year to choose one, while Jack Clark told the IMF and World Bank that the singularity has begun.

DeepSeek raised 15 billion dollars two days after Homeland Security banned its models from federal networks. Microsoft’s MAI-DxO beat physicians on 304 New England Journal of Medicine cases and on routine emergency room visits.

OpenAI neared a 30 billion dollar round at a 1.25 trillion dollar valuation led by Emirati investors, and Meta created a nonvoting share class to keep funding a buildout of up to 145 billion dollars. Governments in Finland, Germany, and the Netherlands pulled websites down rather than wait. Here are the ten developments that defined the week.

Get Anca Trifan | Tree-Fan Events Productions LLC’s stories in your inbox

Join Medium for free to get updates from this writer.

Remember me for faster sign in

If you prefer to receive the Weekly AI News Briefing directly in your inbox, sign up via the form, so you don’t miss it: SIGN UP FORM LINK.

PART ONE

1. Anthropic Filed to List on Nasdaq as ANT, Reporting 16.3 Billion Dollars of Revenue and an 80.7 Billion Dollar Loss

Anthropic filed its S-1 with the SEC on Monday morning, applying to trade Class A common stock on the Nasdaq Global Select Market under the symbol ANT. Revenue for the first half of 2026 reached 16.29 billion dollars, up from roughly 1.16 billion a year earlier. The net loss across the same six months was 80.7 billion dollars, and almost none of it went out the door as cash.

• About 72 billion dollars of the loss is a non-cash remeasurement of convertible notes and other financing instruments, which gained value as Anthropic’s own valuation climbed after a 65 billion dollar Series H in May priced the company at 965 billion dollars. The operating loss was 693 million dollars, down from 2.14 billion a year earlier, and operating cash flow ran positive at 3.93 billion.
• The filing discloses noncancelable cloud and infrastructure commitments running past 450 billion dollars over seven to ten years, plus one equipment lease carrying about 125.2 billion dollars of rent across five years, and warns that missing certain obligations could trigger defaults elsewhere and cut off access to essential compute.
• Anthropic’s own risk factors state that its technology could pose catastrophic or existential risks to humanity, that a severe AI incident could exceed the insurance industry’s capacity to absorb the damage, and that the FTC issued an investigative demand in October covering consumer protection, privacy, data protection, and product safety.

Why it matters: Goldman Sachs leads a syndicate of 22 banks, shares could trade before Thanksgiving, and the first reactions split along the only line that matters, which is whether the loss is real. Strip out the financing remeasurement and the business ran a 693 million dollar operating loss on 16.29 billion dollars of revenue with positive operating cash flow, the strongest set of numbers any frontier lab has put in front of public investors. The comparison everyone reached for within hours was SpaceX, which raised nearly 86 billion dollars at a 1.77 trillion dollar initial valuation and then traded below its IPO price, closing Friday at 147.95 against a 135 dollar offer. Axios framed the listing as a referendum on the business of AI, prediction markets put the company near 1.9 trillion dollars, and the compute commitments are where a skeptic starts, because that bill is contractual while 80 percent of the revenue is usage-based and can simply stop. The Public Benefit Fund in story four turns from a governance document into transferable money the day this stock is liquid.

2. OpenAI’s Last-Generation Model Found Two Critical PHP Flaws and the NSA Told Operators to Pull Servers Offline

PHP maintainers released emergency patches for two vulnerabilities that let an attacker hijack a server without logging in. Security firm Calif found the first on August 1 using OpenAI’s GPT-5.6 Sol, and Google-owned Wiz independently caught both weeks later with Sol and Gemini 4 Argon. The language runs roughly 70 percent of the web.

• The National Security Agency issued an advisory telling operators who could not patch immediately to take systems offline, and governments in Finland, Germany, and the Netherlands pulled public websites down rather than hold them open through the window.
• OpenAI worked alongside the PHP maintainers for two months before disclosure, calling the pair one of the most impactful security issues in recent memory and saying several more significant vulnerabilities are still working through its disclosure pipeline.
• Cloudflare reported seeing early attacks emerging almost immediately, and researchers tracking exposure counted roughly 1.5 million of 2.6 million scanned hosts still unpatched days after the fix shipped.

Why it matters: Sol is two generations behind the current frontier, so what changed here is the cost of looking rather than the ceiling of what can be found. Pre-AI code was audited by people who were expensive and scarce, and a decade of those audits missed a server takeover sitting in the most widely deployed language on the internet. Defenders got about a two-month head start on this one, which is the entire margin the security model now rests on. The hacker campaign in story eight shows what the same price drop buys on the other side.

3. JPMorgan, Bank of America, and Morgan Stanley Began Shutting Third-Party AI Agents Out of Customer Accounts

Axios reported on October 22 that major U.S. banks found third-party AI agents logged into customer accounts using the customers’ own credentials, with some calling support lines or taking actions nobody had requested. One executive described the situation as horrifying. Morgan Stanley moved to block third-party agents outright.

• JPMorgan, Bank of America, and Citi chose restriction over a full block, with most institutions in the reporting pushing agents toward read-only access while they work out what else to allow.
• Most observed cases involved agents the customer had authorized, though one bank executive told Axios the institution could not reliably separate a sanctioned agent from a rogue one.
• The stated fear is hallucination with money attached, an agent granted purchase or transfer authority acting confidently on a misread instruction.

Why it matters: The complaint here is about identity rather than capability, and no amount of additional model quality resolves it. A bank can verify a credential and cannot verify a mandate, which leaves read-only as the only setting it can defend to a regulator. Anthropic launched Claude Duet the same week requiring human sign-off on every payment, signature, and outgoing message, which reads as a direct answer to exactly this objection. Whoever builds the registry that tells an institution which agent may do what will set the terms for consumer AI in finance.

4. Anthropic Set Aside Five Percent of Its Equity for a Public Benefit Fund

Anthropic introduced the Anthropic Public Benefit Fund, a company-owned LLC that would receive 5 percent of its equity if shareholders approve, worth roughly 48 billion dollars at the May valuation. The structure is designed to grow as the company does.

• Each 50 percent rise in Anthropic’s value above an unannounced baseline can add another 2 percent at the board’s discretion, a schedule the company framed explicitly against the escalating pay packages going to technology executives.
• The fund can back nonprofits or make direct distributions to specific populations, with workers displaced by AI named as one target and no equity permitted to go toward lobbying or political spending.
• About 20 percent of Anthropic’s shares now carry some public-benefit designation, counting the new fund, the co-founders’ pledge to give away 80 percent of their wealth, and the stock the company matches when employees donate.

Why it matters: Equity pledges cost nothing until someone liquidates them, and the S-1 filed eight days later puts a date on that. The timing does its own work: the same company published hiring data this week showing its junior technical pipeline collapsing, which makes a fund aimed at displaced workers read less like philanthropy and more like an actuarial estimate. Elon Musk answered within a day by promising a SpaceX share to every American child through Trump Accounts. The competition over who compensates the public has opened well before the disruption has finished arriving.

5. Altman Gave OpenAI Staff Three Endgames for the Safety Race and a Year to Pick One

The Information reported that Sam Altman delivered an internal talk laying out three ways the AI safety race could end. He called the current approach, shipping models once they clear safety tests and flagging concerns, pacing 1.0. The successor decision is due within a year.

• Altman’s stated preference is a global enforcement body modeled on the International Atomic Energy Agency, a structure OpenAI has lobbied for since 2023 and which would require agreement from labs it competes with.
• The second option, which he called the Demis plan after DeepMind’s Demis Hassabis, folds the leading labs into a single entity serving partner countries, closer in shape to CERN than to a company.
• The third, named for Eliezer Yudkowsky, has one lab riding self-improving AI far enough ahead to melt everyone else’s GPUs, an outcome Altman called terrible.

Why it matters: Every one of the three paths depends on rivals cooperating, and the only instrument available for that is persuasion. Altman’s condition that pacing 2.0 include a speed limit for self-improving systems landed within a day of Jack Clark saying Dario Amodei has floated a ceiling on intelligence, which puts the two largest American labs in agreement that a brake is needed and in no agreement about whose hand is on it. DeepSeek closed 15 billion dollars the same week, and Anthropic is about to answer to public shareholders. The plan where everyone gathers under one roof gets less plausible with every quarter that capital reaches labs outside the room.

6. Anthropic’s Own Hiring Data Showed Junior Technical Roles Falling From 27 Percent to 9 Percent

Anthropic published a study of how AI has reshaped its own workforce, and the hiring figures are the part that will travel furthest. Junior roles dropped from 27 percent of technical hires to 9 percent. By August, employees’ agents were logging six times the working hours of the people directing them.

• Engineers now ship five times the code changes they did six months earlier, and Anthropic found agent-written code is deleted or rolled back no more often than human-written code was.
• About a third of the company’s machine learning managers said AI had them hiring fewer people, and employees past a year of tenure switch departments 50 percent more often than in 2025.
• Claude agents sit idle roughly 80 percent of the time during interactive sessions, waiting on a human to reply, which makes the person the slowest component in the loop.

Why it matters: A company selling the technology has every incentive to report the opposite finding, which is what makes its own entry-level numbers worth more than any survey. The 80 percent idle figure carries further than the hiring cut: when the binding constraint is how fast a human can read and approve, the engineering response is to take the human out of that step, which is precisely the autonomous-researcher direction every frontier lab is pointed. The S-1 names the same problem in its risk factors, warning that human reviewers may be unable to keep pace with autonomous systems, which is a remarkable thing to tell prospective shareholders about your own product.

PART TWO: Worldwide AI News

7. DeepSeek Raised 15 Billion Dollars Two Days After Homeland Security Banned It From Federal Networks

DeepSeek closed roughly 15 billion dollars, double its own target, in one of the largest funding rounds in Chinese technology history. Bloomberg reported Tencent and battery manufacturer CATL among the backers. The round landed two days after the U.S. ordered federal agencies to strip the lab’s models out of their systems.

• The Department of Homeland Security formally banned DeepSeek across civilian federal networks, giving departments 60 days to request exceptions and extending the prohibition to third-party tools built on the lab’s models.
• DeepSeek open-sourced DeepInfer the same week, the serving engine behind its V4.1-Flash model, with Huawei chip support that lets companies run it in production without Nvidia hardware.
• V4.1-Flash is also one of the two models behind this week’s 100-company intrusion campaign, running alongside Claude Code in NorForge’s account of the attacks.

Why it matters: A federal ban followed two days later by a record round demonstrates that American procurement no longer sets the price of a Chinese model. The Huawei support matters more than the capital: a serving stack with no Nvidia dependency removes the chokepoint the export controls were built around, and it does so in open source where no license can be revoked. Washington can keep DeepSeek off its own networks. The rest of the deployment surface is now out of reach of that instrument, including the one in story seven.

8. One Operator Used AI Agents to Break Into More Than 100 Companies at About 50 Dollars Each

NorForge found more than 20 hacker control panels sitting on the open internet with no password protection, then watched the campaign they were running. A single Chinese-speaking operator had pointed AI agents at over 1,200 websites and breached more than 100 companies. Reuters identified Cummins, Deere, and SharkNinja among the victims.

• The operator’s instructions ran under 100 Chinese characters, handed to agents on the open-source Cairn framework along with a pre-screened list of high-traffic online stores.
• NorForge priced a successful break-in at roughly 50 dollars in model costs, with a typical one taking close to a day on Claude Code and DeepSeek V4.1-Flash.
• SharkNinja called the threat contained, Deere cited remediation actions, and Cummins said it was investigating, with no public confirmation yet of what data left those networks.

Why it matters: Reaching a thousand targets used to require a team, and that labor requirement is what kept most companies below the threshold of being worth attacking. One person with an estimated 5,000 dollars in tokens now covers the whole operation, which inverts the economics defenders have planned around for two decades. The banks restricting agents this week are reacting to the authorized version of this capability, the one that arrives with permission and a logged-in session. This version asks for nothing and waits for no framework.

9. Jack Clark Told the IMF and World Bank That the Singularity Has Begun

Anthropic co-founder Jack Clark told economists and policymakers at the IMF and World Bank annual meetings that the singularity, the point where AI accelerates discovery faster than people can follow, has now begun. He brought benchmark movement as evidence. Scores on Humanity’s Last Exam went from 8 percent to 67.7 percent since 2025, and FrontierMath from 2 percent to 94 percent.

• One Anthropic model projects U.S. GDP at 44 trillion dollars by 2030 under fast diffusion, a figure Clark conceded sounds outlandish and paired with a warning of very disruptive effects on jobs.
• Clark described the world’s 10 million scientists soon directing a billion virtual colleagues, and urged governments to build physical labs and robots before the bottleneck moves from thinking to doing.
• He said Dario Amodei has floated limits on the ceiling of intelligence, and that there is only a small number of years left to act on safety.

Why it matters: Clark has made the acceleration case before, in blog posts and at conferences where the audience already agreed. Saying it to the institutions that write growth forecasts and lend to finance ministries moves the claim out of lab discourse and into fiscal planning, where someone has to either budget against it or publicly discount it. His intelligence-cap language and Altman’s speed-limit language surfaced within a day of each other, from rivals, with no coordination between them. Two competitors arguing for a brake that each would have to apply voluntarily is the clearest available measure of how little machinery exists to apply one.

10. Microsoft’s MAI-DxO Beat Physicians on 304 NEJM Cases and on Routine ER Visits

Microsoft published initial results for MAI-DxO, a diagnostic system that runs AI agents as a virtual panel of doctors proposing diagnoses, ordering tests, and challenging each other as results come back. It outperformed physicians on hard cases and on ordinary ones, while ordering fewer and cheaper tests.

• On 304 complex New England Journal of Medicine cases it more than doubled the accuracy of physicians working with reference tools, and on typical emergency room presentations it reached 80 percent against 47 percent.
• One configuration hit 81 percent on the NEJM set at less than half the cost of comparable systems, with the gains holding when newer models including GPT-5.5 were substituted in.
• Beth Israel Deaconess in Boston will host the first real clinical test, measuring the system’s diagnoses against clinicians’ assessments on live patients.

Why it matters: Diagnosis is the piece of medicine that most closely resembles a benchmark, which is why it fell first and why these numbers prove less than they appear to. Nothing in the study tracks a patient over months, revises a treatment that stopped working, or sits in a room while someone decides what they are willing to live with. The Beth Israel trial is the one that will count, and it puts a hospital in the position the banks occupied this week, deciding how much authority to delegate to a system that performs well and cannot be meaningfully asked why.

The Week’s Throughline

The cheapest thing in AI this week was looking at code. Calif pointed a model two generations off the frontier at PHP and found a server takeover that a decade of human auditing had walked past, and the NSA responded by telling operators to unplug. Three days earlier, one operator pointed the same class of tool at 1,200 websites and got into more than 100 companies for roughly 50 dollars each. The capability is identical in both stories and the difference is entirely who ran it first. Defenders held a two-month lead on PHP. Nothing in the economics guarantees a second one.

The institutions that have to absorb that are reaching for the same lever. Banks caught agents inside customer accounts and cut them back to read-only, not because the agents reasoned poorly but because a credential proves who logged in and says nothing about what they were authorized to do. Anthropic shipped Duet the same week with human sign-off required on every payment and message. Beth Israel Deaconess agreed to test a diagnostic system that beats its own physicians and cannot explain itself, which is the same delegation question wearing a white coat. Every one of these decisions is about how much authority to hand a system whose competence is measurable and whose mandate is not.

Meanwhile the money kept moving in the direction nobody is governing. DeepSeek doubled its target to 15 billion dollars two days after Homeland Security banned it from federal networks, then open-sourced a serving stack that runs on Huawei silicon and needs no Nvidia at all. OpenAI neared 30 billion dollars from Emirati investors at a 1.25 trillion dollar valuation, Meta engineered a nonvoting share class to fund 145 billion dollars without losing control, and Anthropic set aside 5 percent of itself for the public while reporting that its own junior hiring had fallen to 9 percent. Altman spent the week describing three endings that all require rivals to cooperate, and Clark told the IMF the acceleration has already started and that a few years remain. Both men are asking for a brake. Neither of them, and no institution that met this week, can say whose hand would be on it.

Briefing compiled on October 26, 2026. Sources include Axios, CNBC, Reuters, Bloomberg, Fortune, TechCrunch, The Information, DeepLearning.AI’s The Batch, Every, NorForge, OpenAI, Microsoft AI, and Anthropic. All stories verified against original reporting.

If your team is using AI but still wasting hours searching through inboxes, spreadsheets, CRMs, proposals, contracts, notes, and event files, book an AI Workflow Intelligence Audit. We’ll map one messy workflow, find the leaks, and decide what should be automated, connected, or rebuilt first. Use this booking link, or add “AUDIT” in the comment section.