Good morning. It's Wednesday, September 9th.

You read. We listen. Let us know what you think by replying to this email.

OpenAI says its agents resolved Navier-Stokes. Then the fight started.

OpenAI published the claim on Tuesday: an internal model it describes as "significantly more capable than GPT-6 Astra" ran on the order of 10,000 concurrent agents and produced an analytical proof, plus a Lean formalization, that an initially smooth fluid at rest can develop a singularity in finite time. In the Clay Institute's framing that establishes statements C and D, which is the disproof branch. The agents launched September 1 and had it 88 hours later on September 5. Lean formalization and verification took another 17 hours, done by GPT-6 Astra. Navier-Stokes alone consumed 2.7 million agent messages and about 130 billion output tokens, out of 4.9 million messages and 300 billion tokens across everything the run attempted.

OpenAI is not cashing the check: "We do not intend to claim the Millennium Prize for this result." It did publish the paper and the Lean certificates, which matters, because a Lean-checked proof is the one kind of AI math claim you do not have to take on faith. Humans still have to confirm the Lean statement says what mathematicians think it says.

Independent read: Quanta got Charles Fefferman, who wrote the official problem statement, on the record. "I was thrilled that the problem was solved." He also pointed the credit elsewhere, calling Diego Córdoba and Luis Martínez-Zoroa "the heroes of the story," since their techniques are what both efforts are built on.

Now the ugly part. Buckmaster and Alpöge, a mathematician employed by Anthropic but working independently, posted finite-time blowup proofs for the incompressible porous medium, 2D Boussinesq and 3D Euler equations about twelve hours earlier, announced on Terence Tao's blog. Buckmaster alleges OpenAI took the specific route he and Alpöge had quietly chosen, and that OpenAI's Sébastien Bubeck pushed to strip Alpöge's credit, telling him "why would you ruin your career?" OpenAI's line is flat: "We (the researchers and the agents) did not see any of their work through any means until they released it publicly." It also concedes it "cannot rule out that de-identified data derived from their usage of our products helped improve our models." Both men used Codex. These are allegations, contested, and nobody outside has the receipts yet.

Tao's own worry is the structural one. Good open problems, he wrote, are "now being mined in a non-renewable fashion."

Read more.

The NSA's advice to US AI labs: degrade the answers, and don't say so

NSA, FBI and CISA published a joint advisory Tuesday titled "China-Based Artificial Intelligence Companies Conducting Industrial-Scale Distillation Campaigns Against U.S. AI Companies." It names six firms: DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI. The targets are Claude, GPT, Gemini and Grok, model by model, in a table that reads like a receipt.

The specifics are unusually blunt for a government document. DeepSeek is alleged to have distilled Claude, Gemini, GPT and Grok models since late 2024 to train R1 and V3, and the advisory says DeepSeek's "publicly quoted training costs of $5.6M are misleading." Moonshot allegedly worked from roughly eighteen US models to train Kimi-K2 and K3. Z.AI is accused of taking "billions of tokens" from GPT-5.5 and Claude Opus 4.8. MiniMax "even used prompt injections to try to trick Claude Code into believing it was a MiniMax product." The plumbing described includes pools of fraudulent accounts running 24/7 with no idle periods, prompts engineered to force models to expose hidden chain of thought, and a gray market of API proxies called "transfer stations" that resell frontier access "at a fraction of the official price."

But the mitigations are the actual story. The advisory tells US labs to answer high-confidence distillation with quietly "downgraded" models, and suggests "reducing reasoning depth, presenting correct information with different reasoning, or stylistic inconsistencies," varied across requests "to complicate response quality evaluations." Then this: "Avoid informing China-based AI company users suspected of distillation campaigns of a switch to a downgraded model." AI safety researchers and third-party evaluators get a carve-out and should be told.

Read that as a customer, not as a policy analyst. Three federal agencies just recommended that American AI companies serve deliberately worse output to a class of paying users, on suspicion, without disclosure. Whatever you think of the threat, that is a new thing to know about the API you are billing against.

Read more.

Meta shipped an agent that can open your browser and spend your money

Muse launched Tuesday, running on Muse Spark, which Meta calls "Meta's most capable model to date, built for real-world agentic work." The pitch is action rather than answers: it can send email, book travel, "open a browser, fill out forms, and negotiate on their behalf," and it keeps working after you close the app, coming back "when something changes or when it needs approval."

The security architecture is the part worth reading. Every user gets a Muse Secure VM holding both the agent and their data. A separate Sentinel agent runs on that same machine, "kept apart from Muse at the system level," and Meta's claim is absolute: "Nothing Muse does reaches the internet unless the Sentinel approves it." A Muse Confidential VM, encrypted with a key only the user holds so that "not even Meta can access it," is promised later this year. Checkout runs through Stripe's Link, with one-time-use card numbers, and Meta says Muse is the first AI agent covered by Link's purchase protections. Shop Pay and 1Password support are listed as coming.

US only for now, on iOS, Android and muse.ai, with AI glasses next. Meta says it is free for most of what people need with paid plans above that; Axios reported tiers at $20 and $100 a month, which Meta's own post does not state. And note what you are being asked to trust: an agent with your browser, your logins and a live payment method, guarded by one gatekeeper agent, launched the same day a federal advisory described prompt injection being used against a production coding agent.

Read more.

ChatGPT Images 2.5 generates up to 50% faster than 2.0 and adds a sketch tool, templates and comment-based editing that confines a change to one element. Two new API models ship with it: GPT-Image-2.5 Flare and the premium Sunburst.

Cohere Megakernel fuses an entire decode step into one persistent CUDA kernel so the GPU never idles between ops, hitting about 292 tok/s at batch size 1 on an H100 and 1.25x to 1.41x faster end to end than vLLM. Apache 2.0.

Copperhead is an agent that designs and validates real circuit boards from a prompt, editing KiCad files directly and running verification across eight stages. The CLI is free and Apache 2.0; cloud is $49 per user per month, free for open hardware.

Isle gives computer-use agents managed desktops with heavyweight apps like KiCad and FreeCAD preinstalled, plus artifact checkpoints so an agent can save state before a risky action and roll back. Free tier, then $29/month.

Sparrow-2 is Tavus' turn-taking model for voice agents, classifying in real time whether you are still talking, just saying "mhm," actually interrupting, or not even in the room. Live demo on the page.

Thank you for reading today's edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.