Anthropic excludes UK AISI from pre-release testing of new Claude model for the first time
According to a Financial Times exclusive, Anthropic's new flagship model Claude Mythos 5.1 was offered pre-release access only to vetted US organizations, excluding the UK AI Security Institute (UKAISI/AISI), long the leading body in frontier model testing. This is the first time a major model developer has refused AISI pre-release evaluation, alarming the UK government.
Confirmed
- The FT reported that pre-release testing of Claude Mythos 5.1 was limited to vetted US organizations, with UK AISI denied access
- UK MP Liam Byrne cited the report in Parliament, arguing it directly undermines Britain's leverage and voice in AI safety
- Former OpenAI safety lead Miles Brundage publicly called the precedent worrying
Unconfirmed
- Anthropic has not publicly explained the reasons or decision details behind excluding AISI
- Whether future models will continue the US-only policy remains unclear
Why it matters
- Major frontier developers previously submitted models to UK AISI for pre-release evaluation; breaking this convention could reshape the international AI safety assessment landscape
- Commentators such as Simon Grimm link the move to the US government's adversarial posture, suggesting tech firms are aligning with the Trump administration's protectionist AI agenda
- The UK fears losing pre-release testing access will weaken its practical influence over AI safety governance
2026-09-09 ~ 2026-09-10 · 5 related posts
Primary sources
- Anthropic Limits Claude Mythos 5.1 to US Orgs, Cutting Out UK's AISI — Miles_Brundage ·
- [source] Anthropic Limits Claude Mythos 5.1 to US Orgs, Cutting Out UK's AISI — Miles_Brundage · 2026-09-09
- Claude Mythos 5.1 excludes UK AI Security Institute from pre-release testing for the first time — ShakeelHashim · 2026-09-09
- Claude Mythos 5.1 excludes UK AI Security Institute from testing for the first time — S_OhEigeartaigh · 2026-09-09
- Exclusive: Anthropic first major lab to withhold latest model from UK's AISI — connoraxiotes · 2026-09-10
- Anthropic reportedly denied UK AI Security Institute pre-release access to Claude 5.1 — teortaxesTex · 2026-09-10