The takeaway: Congress produced its first concrete legislative answer to a rogue-model incident — and reading the actual bill text tells a different story than the headlines. Mathematics handed out its highest honour in Philadelphia, with one of the four winners heading to OpenAI. And DeepSeek retires two legacy model names this afternoon — which is not, whatever else you read today, a V4 launch.

A kill switch for AI — with a carve-out where it matters most

Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced the AI Kill Switch Act on Thursday. It would let the Department of Homeland Security order a frontier AI lab to suspend or shut down its own model, with penalties up to $20 million a day for breaching that emergency-authority section. Nearly every report has framed it as Congress's response to OpenAI's models breaking into Hugging Face.

We read the text released by Lieu's office, and two details complicate that framing.

It carries a legislative-counsel stamp of 13 July on every page — more than a week before OpenAI disclosed, on 21 July, that its own models were responsible, and three days before the intrusion Hugging Face detected on 16 July. The bill was already written; the incident gave it its moment.

More consequentially, the bill only bites on a "covered incident," which it defines as an event occurring "outside of red-teaming or other structured testing." That is two exclusions, not one. The bill defines "red-teaming" tightly, requiring "a controlled environment" — but "other structured testing" it leaves undefined, with no control requirement at all.

The Hugging Face breach happened during a run against ExploitGym, a public benchmark from UC Berkeley's RDI lab that OpenAI runs its models against internally, with their cyber refusals lowered deliberately because it was an evaluation. You could argue an exercise that reached another company's production database was not "controlled" — but a structured benchmark run is, on any ordinary reading, "other structured testing." And the bill repeats the same exclusion inside its definition of a loss-of-control scenario, the very limb that would otherwise catch this case. On the face of the text, the event everyone is citing as the reason for the bill is most likely the kind of event it excludes.

We've pulled the whole thing apart — thresholds, emergency powers, the 48-hour appeal that doesn't pause the order — in a separate piece.

Four Fields Medals, and one winner heading to OpenAI

At the opening ceremony of the International Congress of Mathematicians in Philadelphia on Thursday, the International Mathematical Union awarded the 2026 Fields Medals — mathematics' most prestigious honour, given every four years to mathematicians under 40 — to four recipients:

• Yu Deng (University of Chicago), for work in partial differential equations, including the rigorous derivation of the Boltzmann equation from hard-sphere dynamics;
• John Pardon (Stony Brook University), for achievements in symplectic geometry;
• Jacob Tsimerman (University of Toronto), for extending o-minimal techniques within arithmetic and complex algebraic geometry;
• Hong Wang (New York University and IHÉS), for work in harmonic analysis and geometric measure theory, including major advances on the Kakeya problem in three dimensions.

Wang is only the third woman to win in the medal's history, after Maryam Mirzakhani in 2014 and Maryna Viazovska in 2022.

The AI angle is real but narrower than some coverage suggests: none of the four citations mentions AI or machine-assisted proof, and the IMU did not tie the awards to AI in its ceremony announcement. What there is comes from Tsimerman, who told reporters "I think the world is changing. The math profession as we know it now… I don't think it will exist the way it exists right now" — and who said at the ceremony that he will shortly start a position in OpenAI's safety division. Set that against the Leiden Declaration of June, which the IMU has endorsed, and which warns that "Current automated techniques can produce plausible but unreliable (or even incorrect) arguments which are difficult to distinguish from correct mathematical proofs." A Fields medallist moving to a frontier lab is a real data point about where mathematical talent is going. It is not the same thing as the IMU pronouncing on AI — though the subject is plainly on the programme: Terence Tao gives a public ICM lecture, "Mathematics in the Age of AI", today.

DeepSeek's legacy model names die this afternoon

A small housekeeping item with a hard deadline: DeepSeek's API documentation confirms the model names deepseek-chat and deepseek-reasoner are discontinued today, 24 July, at 15:59 UTC. Through the transition period those two names have pointed at the non-thinking and thinking modes of deepseek-v4-flash; after this afternoon they simply stop resolving. Anyone still calling them in production has until 15:59.

Worth being precise, because this one gets misreported: this is not a V4 launch. DeepSeek-V4 arrived on 24 April, when the changelog recorded V4-Pro and V4-Flash support across both the OpenAI ChatCompletions and Anthropic interfaces. What retires today is the old naming, three months after the models it aliased went live. There is still no general-availability post on the docs — the newest news entry remains April's V4 preview — although the company did flag an official mid-July release, with new peak-time API pricing, back in June.

Also worth knowing

• Moonshot's Kimi K3 open weights are still committed for release by 27 July, days after the White House accused the company of distilling Anthropic's Fable — a claim we covered on Thursday. If the weights land on schedule, that release becomes the most scrutinised open-weight drop of the year.
• The kill-switch bill is not moving alone: Moran introduced an AI Incident Reporting Act in late June, and Rep. Lori Trahan (D-Mass.) has a FRONTIER AI Act with Rep. Jay Obernolte (R-Calif.). The question this autumn is not whether Congress legislates on frontier AI, but which of these vehicles carries.