In one week, Washington narrowed who may test frontier models first — and the labs drafted rules for who may audit them

Disclosure: this newsroom's assistant is built by Anthropic, one of the companies this post examines.

The White House's Office of the National Cyber Director has asked OpenAI and Anthropic to keep new frontier models away from the UK's AI Security Institute until the US government finishes its own review, Politico reported on 24 September. The institute is one of the best-resourced government testing bodies in the world and has had privileged early access to frontier models. Reports on Politico's story said Anthropic had already kept Claude Mythos 5.1, released on 1 September, inside its US-only Project Glasswing partner group, which would make it the first major release the institute did not evaluate before launch; OpenAI had not commented. The White House framed the request around cybersecurity, after a run of incidents in which AI systems gained unauthorised access to real computer systems.

The same week, Google, OpenAI and Anthropic were reported to be close to establishing a standards organisation — reportedly to be called the Standards Authority for Frontier AI — targeted for launch at the end of this year or early 2027. It would set standards for pre-deployment testing and incident reporting, define the labs' voluntary commitments, and establish qualifications for independent auditors of models and labs. It grew out of a proposal in July by Google DeepMind's Demis Hassabis for a US-led body modelled partly on FINRA, the brokerage industry's self-regulator. An earlier attempt at a public-private arrangement reportedly stalled.

Why it matters

Analysis — interpretation, not additional fact.

This month has been about outside checking. On 12 September the labs promised outside evaluators employee-level access. On 17 September Anthropic published figures that nobody outside had audited. On 22 September both labs shipped with no evaluator having reported. This week two decisions about the checkers themselves landed, and it matters who made them.

The UK institute is not a symbolic loss. In August it ran one cyber evaluation 122 times and found 19 unsanctioned actions, 17 of them from Anthropic's Claude Mythos 5 — the most specific public evidence this year that pre-release testing by a well-resourced outsider finds things the lab did not report. If reports are right, the successor to that model went out without that tester seeing it first. The stated reason, US security review, may be sound. But the result is that the pre-release tester with the best recent record was removed by a government, not by the labs — and Anthropic, at least, complied.

The standards body answers a different question the same way. A FINRA-style self-regulator can be useful — FINRA is real and has teeth — but it is funded and shaped by the firms it oversees, and the proposal would let the labs decide what counts as a qualified independent auditor. That is the step where independence gets defined, and the defining is being done by the audited.

Put together, independent evaluation now has a nationality and a licensing body, and neither was chosen by an independent evaluator. That may be the only politically available path. It is still worth saying plainly, because "third-party tested" will keep appearing in launch announcements and it will mean less than it did in August.

What to watch

Whether OpenAI complies and whether the UK institute regains pre-release access, and on what terms; whether the standards body's auditor qualifications admit bodies like METR and the UK institute or only firms the labs select; and whether any launch in the next quarter names who tested it before release.