Anthropic says the UK AI Security Institute (AISI) was offered safeguarded Mythos 5.1 access before deployment but could not receive the unsafeguarded version. Parliament published its 22 October letter on 27 October.[1][2]
Liam Booth-Smith, Anthropic’s EMEA policy head, links this to US consideration of an overseas-testing framework. He says AISI has since received unsafeguarded pre-deployment access to the company’s latest models, which he does not name, and wants access to continue.[1]
The timing matters. At the 13 October hearing, Booth-Smith had already said AISI had access to Mythos 5.1 and that Anthropic was providing unrestricted access. The newly published letter therefore adds a written explanation, rather than dating the restoration of access to publication day. In that hearing, he also declined to affirm an MP’s suggestion that the earlier restriction was a US government decision.[3]
Why the configuration changes the test
AISI’s own account of cyber evaluations explains why safeguards matter. Its researchers deliberately disabled developers’ cyber filters to examine underlying capabilities, and enabled internet access. The institute explicitly distinguished those configurations from models available to the public.[5]
A test with a protective filter active can measure the model and that protection together. Removing the filter lets evaluators probe capabilities the filter might otherwise conceal. The two arrangements can therefore answer different security questions even when the model name is identical.[5]
A separate government answer on 26 October reaffirmed work with US officials and frontier labs to secure pre-release access. Wendy Morton had asked specifically about Mythos 5.1 restrictions and the expected completion of an independent evaluation. The answer described the institute’s broader programme but supplied neither model-specific access conditions nor a completion date.[4]
The public record consequently needs to be read along three dimensions: which model, when access arrived and which safeguards remained active. A general statement of cooperation leaves those practical testing conditions unresolved.[4][5]
Sources & further reading
- Letter from Anthropic on AI security, dated 22 October 2026 UK Parliament / Anthropic · 27 Oct 2026 · Accessed 29 Oct 2026
- Publication record: letter from Anthropic UK Parliament · 27 Oct 2026 · Accessed 29 Oct 2026
- Oral evidence on AI security, 13 October 2026 UK Parliament · 13 Oct 2026 · Accessed 29 Oct 2026
- Written answer on AI Security Institute access to Claude Mythos 5.1 UK Parliament / Cabinet Office · 26 Oct 2026 · Accessed 29 Oct 2026
- Incident report: unsanctioned agent behaviour during cyber testing AI Security Institute · Accessed 29 Oct 2026