Anthropic withholds Mythos 5.1 from UK pre-release testing
Anthropic did not give the UK AI Security Institute pre-release access to Claude Mythos 5.1, launched on 1 September, while vetted US organisations received access. It is the first time the institute has been excluded from a pre-release evaluation of an Anthropic model. The UK Cabinet Office says cooperation continues, but Anthropic has not explained the decision.
Artificial Intelligence··Evening
The UK institute is excluded for the first time
Anthropic did not submit Claude Mythos 5.1, released on 1 September, to the UK AI Security Institute for pre-release evaluation. Vetted organisations in the United States received access instead. ITPro and IBTimes UK report the same access decision and the institute's exclusion. The institute had evaluated Anthropic models before release in the past; this was the first Anthropic model for which it was left out at that stage. Because the decision became public after the model's release, the missed pre-release review could not be restored retrospectively.[1], [2]
The Cabinet Office says cooperation continues
A UK Cabinet Office spokesperson said the institute continues to work closely with industry partners, including Anthropic, to make models safer. The spokesperson also noted that the institute tested OpenAI's GPT-6 Astra before release. That statement does not say the institutional relationship has ended, but it also does not alter the fact that Mythos 5.1 was not provided for this particular pre-release evaluation. Anthropic did not explain the decision and did not respond to ITPro's request for comment. The reports do not establish whether the access difference arose from security, commercial or other considerations.[1]
Access to Mythos was already restricted
Mythos first appeared under restricted access in April. The United States placed temporary export restrictions on Mythos 5 and Fable 5 in June, and the new Mythos 5.1 was also released through an access arrangement limited to vetted organisations. That history shows the model was not broadly distributed, but it does not explain the UK institute's changed evaluation role. The new element reported on 9 September is that some US organisations received access while the UK's publicly backed security institute was excluded from pre-release review for the first time, without an explanation from Anthropic.[1]
