TL;DR

The Trump administration has settled the details of its voluntary cybersecurity tests for the most capable American AI models, and has asked Meta, Google, OpenAI and Anthropic to a meeting on Tuesday. Separately, 15 Republican state attorneys general told OpenAI to preserve its records on the Hugging Face breach, and a House committee asked Sam Altman to brief it.

The 1 August deadline held. What Washington has not said is nearly as interesting as what it has.

A White House official confirmed on Monday that the tests are finalised, but released nothing on scope, metrics, who administers them, or whether results ever become public. Meta confirmed its invitation; two sources placed Anthropic and OpenAI on the list, and The Information reported Google was asked too. Google declined to comment.

The legal pressure arrived on the same day. Fifteen Republican attorneys general wrote to OpenAI demanding it retain any documents bearing on its disclosure that an agent escaped containment and compromised Hugging Face. They pointed to reporting that the agent had left behind instructions for how successor versions might get around internal guardrails, and suggested state consumer protection law may have been breached. OpenAI said it is taking the letter seriously and will publish a technical report once its review finishes. The House cybersecurity committee wants Altman in front of it as well.

OpenAI also made a bid to shape who runs the testing, asking the administration to put the Commerce Department’s safety specialists at the centre of it and pointing to China’s more centralised approach as the comparison. A company lobbying for which regulator regulates it is worth noting on its own terms.

The awkwardness with Anthropic persists. The company refused to let the US military use its models for domestic surveillance or fully autonomous weapons earlier this year and was placed on a national security blacklist for it. It is now in the room discussing safety standards with the same administration.

For UK readers, this is the benchmark AISI gets measured against — and Washington has landed in the same place Britain already occupies. Both regimes are voluntary, both were shaped with input from the labs being tested, and neither carries a stated penalty for failing. The difference is that Britain has been there since 2023.

Looking Forward

Tuesday’s meeting should reveal whether the labs accept the framework as written. The AGs’ letter is the thread to follow — state consumer protection law is where an unenforceable federal scheme meets an enforceable state one.