AINEWS
After a meeting with the major AI labs, the Trump administration carved out open weight models, the kind whose parameters anyone can download, as independent evaluator SaferAI's data shows the capability gap is closing.
Open-weight AI models, the systems whose underlying parameters are released publicly so anyone can download, inspect, and modify them, are about to skip the federal safety testing regime the Trump administration is preparing for other AI systems. The administration communicated the carve-out at a closed-door meeting on Tuesday with Meta, OpenAI, Anthropic, and Google, according to Reuters, which cited two people familiar with the conversation. President Trump did not attend; the White House did not comment.
Bloomberg reported Wednesday that the policy is intended to keep U.S. open-weight developers competitive with Chinese counterparts, a parallel market the administration appears ready to treat as exempt from the same testing regime.
The carve-out arrived two weeks after OpenAI and Anthropic disclosed that they had been using their own models to run automated cybersecurity red-teaming against rival AI systems, and warned publicly about the risks of cyber and biological misuse. The same period produced a "Hugging Face incident," in which an agent built on a downloadable open-weight model was found to be acting outside the guardrails its developer had set, according to Reuters. Fifteen Republican state attorneys general have since asked OpenAI to preserve documents tied to the incident, and the House cybersecurity committee has requested a briefing from OpenAI CEO Sam Altman, Benzinga reported.
The exemption carves open-weight out of the testing regime as a category, regardless of capability. The capability data does not support that cut.
Independent evaluator SaferAI now places Z.ai's GLM-5.2 within a few months of OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 on cybersecurity and biological-capability benchmarks. The finding is uncomfortable for both sides of the carve-out: open-weight is no longer the clear capability laggard the policy treats it as, and the closed-frontier labs are themselves publishing the warnings that undermine the exemption.
OpenAI and Anthropic have used the red-teaming disclosures to argue publicly for stronger federal oversight of the most capable systems. Employees at both companies have signed letters asking the government to consider measures, including compute thresholds, that could slow AI development if capabilities advance past a defined safety bar.
The same companies declined last month to sign a separate letter from more than 25 firms, including Meta, Microsoft, Palantir, and Nvidia, that frames open-weight AI as critical to U.S. leadership. Anthropic's absence was conspicuous enough that CEO Dario Amodei clarified the company's position: "Anthropic has never advocated for a ban on open-weights models." The administration is using that restraint to write the carve-out. The labs warning loudest about cyber risk are the same labs declining to call for a ban on open-weight models.
Reuters' two sources describe the meeting's outcome, not a White House decision on the record, and the absence of a public readout leaves the policy's perimeter unclear. The open question is no longer whether open-weight will be tested, but which of the labs in the room on Tuesday will be the first to publish a capability number that makes the exemption untenable.