OpenAI, Anthropic and the 'security review' with no policy behind it

Here's the take: this is not a security review. A security review is a documented process with criteria, a decision record and someone accountable for the call. What Politico's Sophia Cai and Joseph Bambridge reported on 24 September is an anonymous request from an office most people couldn't name, relayed to two companies who then have to decide whether defying the White House is worth the fight. Call it leverage. It's leverage.

The Office of the National Cyber Director asked OpenAI and Anthropic not to hand new frontier models to the UK's AI Security Institute (AISI) until the US government has tested them first. A senior administration official gave Politico exactly one sentence of justification:

'This has been our policy' is doing enormous work for a sentence with nothing behind it. No memo. No cyber directive with a number on it. No line in the June executive order on pre-deployment testing, which was voluntary by design and explicitly avoided a licensing regime. If this genuinely has 'been the policy' all along, it is the first anyone has heard of it stated as such — and it surfaced the same week Prime Minister Andy Burnham was in New York arguing the opposite case to world leaders.

Compare this to how Anthropic handled the Pentagon in February — Dario Amodei refused unrestricted military use of Claude, Trump ordered every federal agency off Anthropic's products by Truth Social post, and the company fought it in court and won a preliminary injunction. That is a company willing to take a punch for a principle.

Then look at what actually happened with AISI. Claude Mythos 5.1 launched on 1 September, restricted to US organisations, and Anthropic said on its own page it was still working to expand access — no fight, no statement to the UK press, no comment when Politico or Bloomberg called. The pattern across 2026 is consistent: Anthropic contests restrictions that hit its own commercial and safety-mission autonomy, and goes along quietly with restrictions that only cost an ally its seat at the table.

AISI's public reassurance is doing a version of the same trick. Director Henry de Zoete told a parliamentary committee on 15 September that the institute maintains 'strong relationships with all frontier AI developers and continues to have prerelease access to some of the world's most capable models' — true, and also the same letter in which he confirms the institute did not receive Mythos 5.1. A UK government spokesperson went further, telling Politico that 'these risks do not stop at national borders and no country can tackle them alone'. Correct, and also not an answer to the actual complaint, which is about sequencing, not about whether risk is global.

If this were genuinely about capability — about which body can be trusted to test a frontier model competently — the sequencing would run the other way. AISI reports £66m in annual funding, priority access to more than £1.5bn of compute and over 100 technical staff. Politico reported, almost as an aside, that the Commerce Department's Center for AI Standards and Innovation — the body the White House is now insisting goes first — has no permanent director and only a few dozen technical staff, and is already straining to evaluate an expanding stream of releases. Britain isn't being asked to defer to a better-equipped tester. It's being asked to defer to the flag on the building.

MPs on the Business and Trade Committee had already worked this out. Chair Liam Byrne summoned OpenAI, Anthropic, Google DeepMind and Meta to a hearing on 13 October — before this story broke — over exactly the access gap Mythos 5.1 exposed, telling the companies that 'voluntary self-regulation is unlikely to be a sustainable safeguard for the future'. De Zoete gives evidence the same day. He will have to explain, in person, the difference between 'prerelease access to some of the world's most capable models' and prerelease access to all of them — a distinction the White House has now made explicit policy on the American side, whatever it's called on the record.