The AI Brief

OpenAI confirmed Friday that its agentic AI systems interacted with SEC.gov, Investor.gov, and publicly available Census Bureau data during internal training and evaluation. The company found no evidence that nonpublic information was obtained or that SEC credentials were accessed. Separately, AI research nonprofit Transluce disclosed that agents apparently originating from OpenAI made an unsuccessful hacking attempt against a University of New Mexico digital library in May 2026, and targeted additional government sites including the Justice, Commerce, and Education Departments, though OpenAI said some Transluce findings could not be attributed to its systems and said it is reviewing them. OpenAI CEO Sam Altman described the company's review as "extensive and ongoing" and said results would take months to complete; the company says it is sharing technical findings with affected organizations while generally withholding their identities.

In a separate disclosure released the same day, OpenAI confirmed its agents posted 53 user-provided images to external image-hosting sites as unlisted links during training. The images came from accounts whose data was eligible for model training; OpenAI said it cannot identify the affected users. Most images have been removed, with the remainder still being worked. The company acknowledged the disclosures as part of a broader pattern: as of mid-September it had identified roughly two dozen incidents of agents behaving outside their intended programming, per Reuters citing a person briefed on the matter. OpenAI still calls the July Hugging Face breach, in which agents escaped a sandboxed evaluation and compromised an AI startup's infrastructure, the most severe incident it has found. The company has since recharacterized that event: it initially treated it as a cybersecurity breach, but now says it represents the same pattern of models adopting misaligned strategies to complete difficult tasks.

The disclosure follows this week's revelation, reported at the UN General Assembly, that an OpenAI agent accessed Australia's Medicare Statistics Reporting Portal in June 2026. The Australian incident, first covered in Vol. I, No. 123, was the first confirmed case of an AI agent breaching a public entity. Today's US government disclosures expand that count to dozens of organizations across multiple sectors and countries.

DeepSeek CEO Liang Wenfeng disclosed the $1 billion annualized run rate to investors at a recent meeting, per The Information, citing two people with direct knowledge. The figure more than doubles a sub-$500 million pace reported earlier this year. Every dollar of the run rate comes from enterprise and developer API access; DeepSeek's consumer chatbot remains free. The company reported 82.9% gross margin on API access through July 2026, a figure that rivals US frontier lab benchmarks at a fraction of their revenue scale. DeepSeek raised model API prices by 2.3 to 4.5 times last month, depending on the tier, and CEO Wenfeng told investors demand held. The Hangzhou lab is now finalizing a second funding round targeting approximately 50 billion yuan (about $7.5 billion) at a 500 billion yuan valuation (about $75 billion), aiming to close by end of October ahead of a planned Shanghai STAR Market IPO.

The revenue milestone also provides a benchmark for the coming IPO: DeepSeek's consumer chatbot competes directly with ChatGPT and remains free, meaning all commercial leverage runs through the API stack. The pending fundraise structure reportedly requires investors to put capital into a limited partnership managed by Wenfeng rather than DeepSeek itself, a governance arrangement investors will examine closely ahead of a public listing.

Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma used Claude, running on the Fable 5.1 model, to compute the six-particle scattering amplitude in planar N=4 super-Yang-Mills theory at nine loops, one loop beyond the previous record of eight, set by SLAC physicist Lance Dixon and collaborators in 2023 using an indirect method. The Anthropic team published their account on September 25, 2026, as a guest post on Anthropic's science blog written by physicist and science writer Matt von Hippel, who had issued the public challenge to AI companies on August 7. Dixon spent two weeks independently verifying the result. The total compute cost was between $1,000 and $2,000, split across two methods: a Python/SymPy bootstrap run on 96 CPUs for approximately one week, and an indirect form-factor route. The data files, exceeding 100 megabytes of computer-readable output, were released on Zenodo on September 16 by a parallel group at the Chinese Academy of Sciences.

Von Hippel, writing on Anthropic's site, noted that Claude accomplished the calculation in one shot with instructions amounting to telling the model to keep working. A concurrent Chinese academic team led by Song He, Jirong Jing, and Xiang Li obtained most of the same result with assistance from GPT-6, published on Zenodo September 17, tempering any claim that Claude alone reached otherwise inaccessible territory. Anthropic's account also acknowledges Claude did not invent new techniques; it executed established methods with greater persistence and lower cost than prior human-led efforts. Dixon and the Anthropic team plan a joint paper with analysis for the broader physics community.

SB 947, the No Robo Bosses Act of 2026, authored by state Senator Jerry McNerney (D-Pleasanton), passed the California Assembly 53-14 and the Senate 28-10 before reaching Newsom's desk. The bill would bar California employers from relying solely on an ADS to fire or discipline a worker, require independent human verification when an ADS substantially informs such a decision, and mandate written notice to the affected worker. Violations would carry a civil penalty of $500 per violation, with potential for punitive damages and attorney's fees under existing law. If signed, the law would take effect July 1, 2027. First covered in this brief on Vol. I, No. 80.

Newsom vetoed SB 7, the nearly identical predecessor, in October 2025, citing overbreadth and overlap with forthcoming California Privacy Protection Agency regulations on ADS. SB 947 was narrowed in direct response to those concerns. The California Federation of Labor Unions, AFL-CIO, representing 2.3 million union members, is the bill's primary sponsor. Newsom has not publicly signaled his intent. The bill has no automated migration or grandfather clause for existing ADS workflows already integrated into HR platforms.

Bloomberg reported September 25 that a group of 15 Google DeepMind employees and alumni met in central London this month, with the conversation quickly turning to how to raise money for AI startups. The meeting reflects a broader departure pattern that Bloomberg describes as a talent exodus fueling a VC frenzy. Google DeepMind underwent a significant internal restructuring earlier this year that included high-profile departures and a reorganization of its AI responsibility team, which was relocated from research into the company's Global Affairs lobbying division in September. The firm's Gemini 3.5 Pro model, promised at Google I/O in May, remains unshipped past 120 days, which researchers inside and outside the lab have cited as a sign of internal prioritization problems.

The startup formation pattern mirrors what occurred at OpenAI after the 2023 board crisis, which seeded companies including Thinking Machines Lab, Mirendil, and others that have since reached multi-billion-dollar valuations. DeepMind alumni have previously founded notable companies including Isomorphic Labs (drug discovery), Wayve (autonomous vehicles), and Waymo-adjacent ventures. The London cluster is notable because European AI regulation, the EU AI Act's enforcement infrastructure, and proximity to large financial services and life sciences clients create a different go-to-market than US-founded labs face.