The AI Brief

OpenAI CEO Sam Altman posted a summary naming four results as the release's biggest: the quasi-Riemann hypothesis (a zero-free half-plane at real part above 7/8, accompanied by a Lean proof), the Unique Games Conjecture (also Lean-verified), the rational Hodge conjecture for CM abelian varieties (no Lean proof yet), and the free group factor isomorphism problem (Lean-verified). The 722 manuscripts are organized into 372 "result families" spanning 16 areas of mathematics, including number theory, theoretical computer science, operator algebras, and mathematical physics. OpenAI says the model was posed roughly 4,000 problems; the 722 manuscripts represent the subset the company judged worth publishing, an 18% yield. Each result used on average three hours of ChatGPT Pro thinking compute.

The release triggered immediate backlash. Multiple mathematicians accused OpenAI of ignoring a closed-door August meeting at which researchers asked the company to follow academic publication norms, and the release comes weeks after prior allegations that OpenAI's math program scooped collaborative work on the Navier-Stokes Millennium Prize problem. The Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study was consulted on how to release the work, not on whether to. The model used has no name, no model card, and no API endpoint. Lean mathematician Levent Alpöge called the quasi-Riemann result "the most significant moment in mathematical history," while others pointed to conduct as disqualifying the reception.

Verification is uneven by design. Many manuscripts include Lean formalizations; others do not, and OpenAI warns that unformalized results may contain errors. Corrections will be recorded as new versions, with earlier versions preserved. OpenAI says it will fund workshops and conferences to help the community understand the results. The repository is licensed Apache 2.0 at github.com/openai/math.

Google signed two contracts with Constellation Energy on October 6. The first is a 20-year power purchase agreement covering 890 MW of nuclear output from upgrades to 11 existing reactors in Illinois, Pennsylvania, and New Jersey, with electricity from the first upgraded plant expected in 2028. The second is a separate long-term supply agreement for 2,700 MW of additional Constellation capacity inside PJM. The two together total 3,590 MW. Constellation will invest more than $4.3 billion in the project, sustaining roughly 4,400 existing jobs and creating approximately 7,200 construction jobs during the build period. Both companies described the deal as a direct response to PJM's proposed "bring your own power" rules for data centers. Google and Constellation also expanded a technology alliance involving Google Cloud and Gemini Enterprise. The deal follows Constellation's earlier agreements with Microsoft (Three Mile Island restart) and Meta (Clinton Clean Energy Center).

Anthropic announced the expanded CVP on October 6, folding Project Glasswing, the invite-only program that gave critical infrastructure organizations access to Claude Mythos, into a single three-tier structure alongside the earlier Cyber Verification Program. The three tiers are: Defense Access, for security operations center work, incident response, malware reverse-engineering, and vulnerability validation; Red Team Access, adding authorized penetration testing; and Specialized Access, the narrowest tier, reserved for organizations authorized to test safety-critical systems such as flight operating systems and power grids, with each Specialized applicant reviewed in coordination with the US government. All three tiers include Claude Opus 5.5, Claude Sonnet 5.5, and Claude Mythos 5.1, plus future models as they ship.

The scale data comes from the retired Glasswing program: partners verified at least 129,000 software vulnerabilities between April and July 2026, with more than 33,000 rated critical or high severity, plus 5,500 more from Anthropic's own open-source scanning through October. Of 5,674 true-positive findings, only 516 have been patched, highlighting the gap between AI-assisted discovery and remediation. On Anthropic's internal CyScenarioBench, Claude Opus 5.5 completed zero of 50 offensive tasks without the program, four at the Defense tier, and 34 at the Red Team tier. A data-retention requirement applies to all tiers; zero-retention enterprise options are described as shipping later this fall.

Disclosure: Claude, which generates this brief, is built by Anthropic.

OpenAI announced on October 5 that it will test a new visual ad format in ChatGPT during image generation, beginning later in October in the US with an initial group of advertisers. The ads display labeled visual content alongside, not inside, the image being generated, and OpenAI states they will not influence ChatGPT's responses. Example screenshots published by the company show a grocery brand's product images with a "Learn more" button appearing while an image request is processed. The format targets the moment when a user is already engaged in a creative task, giving advertisers a "product inspiration" surface. Placement guardrails assess whether a conversation is appropriate for advertising before serving an ad.

The infrastructure buildout alongside the format announcement is notable: OpenAI expanded its measurement partnerships to include Haus, Measured, and WorkMagic for geo-lift experiments, and is developing incrementality testing tools. The company introduced ads earlier in 2026 to support free and low-cost subscription tiers, expanded to India in August, and has now added Sponsored Agents (launched September 17 with Wayfair and Angi as partners). The visual format is the first placement tied to a specific creative workflow rather than a conversational response. OpenAI has not disclosed which ChatGPT subscription plans will see the ads during the US test.

Meta and Sierra co-announced the Personal Agent Protocol on October 6 at Sierra Summit in San Francisco. Sierra co-founders Bret Taylor and Clay Bavor published the announcement; Meta's engineering lead David Singleton posted in parallel on X with a partly different partner list, a detail that reflects the protocol's early governance state. The v0.1 specification has not been published and is expected later in October, followed by design workshops and a reference implementation. Founding partners named by Sierra include Walmart, Shopify, Stripe, Rocket, Genesys, and Instinct. Neither a governing body nor a license has been disclosed.

The protocol sits at the session and authentication layer: an agent discovers what a merchant offers, starts a guest session or signs in to the customer's existing account over OAuth, and the merchant controls whether the agent reaches it via website, MCP or OpenAPI, or the merchant's own agent system. The standard is designed to carry context across channels. Payments and finer-grained permissions are listed as planned extensions. The protocol enters a crowded field: OpenAI and Stripe launched the Agentic Commerce Protocol in September 2025; Visa and Cloudflare launched the Trusted Agent Protocol in October 2025; and Google's Universal Commerce Protocol added Meta, Amazon, Microsoft, and Salesforce to its council in January 2026. PAP governs the identity and authentication step that sits above all of them.