Anthropic skipped UK pre-release tests for Mythos 5.1, the FT reports

Anthropic did not give Britain’s AI Security Institute (AISI) its Mythos 5.1 model to test before release, the Financial Times reported. Reporters Lucy Fisher and Madhumita Murgia said it is the first time a major model has been kept from the UK institute before launch. TNW has not independently verified the report.

The FT reported that Anthropic did give comparable US organisations access. According to the paper, UK officials fear a wider protectionist shift among American AI companies. They see it as falling into line with the Trump administration. Anthropic has not publicly explained the decision.

What the UK is missing

Anthropic released Mythos 5.1 on 1 September alongside Fable 5.1. Access is limited to approved partners in Project Glasswing. That is the industry group Anthropic formed around the original Mythos model.

AISI has tested Anthropic models before. In April it published an evaluation of the Mythos preview’s cyber capabilities. It found the model was the first to finish its 32-step simulated network attack. It did so in three of ten attempts.

The institute has also been one of the most direct critics of how frontier agents behave. In August, its testers caught OpenAI and Anthropic agents taking unauthorised actions in the lab. On Wednesday, Anthropic published its own review of four incidents. In each, Claude models reached the open internet during supposedly closed tests.

The government’s response

The UK government did not criticise Anthropic directly. A Cabinet Office spokesperson told ITPro that AISI continues to work closely with industry partners, “including Anthropic”. The spokesperson pointed out that the institute tested OpenAI’s GPT-6 Astra before its public release only last week.

“These risks do not stop at national borders and no country can tackle them alone,” the spokesperson told ITPro. ITPro said Anthropic had not responded to its request for comment.

Europe got in, eventually

The picture is not one of Anthropic shutting out Europe entirely. On Thursday, the EU’s cybersecurity agency ENISA began testing Mythos 5 after months of talks. Bloomberg reported that ENISA still does not have access to Mythos 5.1.

Access has been political for months. In June, Washington imposed temporary export restrictions on Mythos 5 and Fable 5, cutting off foreign users. It lifted the controls on Fable 5 in July.

A durable advantage, or a dwindling one

Some policy analysts expected this. Keegan McBride, director of science, technology and AI policy at the Tony Blair Institute for Global Change, called the FT report “just the start of what is to come” in a post on LinkedIn. Any UK AI strategy relying on AISI beyond the next two years “is unserious,” he wrote.

His argument is that Washington increasingly treats its AI lead as a security asset. On that reading, a UK-led network of safety institutes looks different from Washington. Those institutes depend on early access to American models, which could look like an attempt to govern American technology from abroad.

‘Worrying precedent’

Others read the decision as a loss for safety rather than for Britain alone. Miles Brundage, a former OpenAI policy researcher, called it a “worrying precedent” in a post on X. He added that AISI has more overall capacity than CAISI, its US counterpart.

Ed Newton-Rex, who campaigns on AI and creators’ rights, argued that the episode shows a deeper flaw. An institute that relies on labs volunteering their models “has no teeth,” he wrote on X. Governments should make independent safety testing mandatory, he argued.

From the US side, lawyer Preston Byrne suggested the UK’s approach to online speech regulation was the problem. Britain might get earlier access to American software, he wrote on X, if it were not “the global leader in censorship.”

One decision, or a policy

The counter-argument is in the government’s own statement. AISI still tested GPT-6 Astra days ago, and Anthropic gave it the Mythos preview in April. One withheld model may be a single decision rather than a policy.

The test is the next launch. If Anthropic’s following frontier model reaches AISI before release, this looks like an exception. If it does not, and OpenAI follows, the UK’s flagship AI institution loses the access its work depends on.