GovernmentAI-TechBusinessScienceSportsEntertainmentGeneral
AI-Tech

Anthropic restricts Claude Mythos testing to vetted US groups

Miles Brundage warned that Britain retains greater evaluation capacity despite exclusion from the developer's newest frontier cyber and biology model.

By Primary News Team

Anthropic restricts Claude Mythos testing to vetted US groups
AI safety researcher Miles Brundage called the exclusion of the British evaluation body a worrying precedent. Source: X
Published9 Sep 2026, 14:44 Last updated9 Sep 2026, 14:44 Sources
Show reference links Marks each sentence drawn from a source or a contributor

Anthropic excluded the British government's evaluation body from pre-release testing for its Claude Mythos 5.1 artificial intelligence system, restricting early access strictly to vetted American organizations.12 The decision leaves the UK AI Safety Institute out of an Anthropic frontier evaluation cycle for the first time, sparking immediate debate among researchers over national technical capacities.1

Addressing the development on September 9, 2026, AI safety researcher Miles Brundage called the exclusion a worrying precedent.2 Brundage noted that the British institute retains greater overall testing capacity than the US Center for AI Safety Institute, though he stated the American center requires more funding, stable leadership and communication rules, and higher salaries.2

Anthropic launched Claude Mythos 5.1 alongside Claude Fable 5.1 on September 1, 2026.13 The two releases share the same underlying model weights, operating under different safety regimes.43 Fable 5.1 is generally available to commercial customers and includes automated classifiers that intercept sensitive requests.54 When those filters flag risky cyber or biological prompts, requests route down to earlier models like Claude Opus 4.8 and Claude Opus 5.53 Mythos 5.1 operates without those filtering classifiers, exposing the raw model to trusted partners.31

Anthropic restricts Claude Mythos testing to vetted US groups
A bar chart compares Claude Mythos Preview's benchmark performance, as discussed in the passage. Source: Wandb

Anthropic provides Mythos access through specialized vetting channels, including the Cyber Verification Program and the Life Sciences Verification Program run alongside the American government.53 The software company also provides access through Project Glasswing, a defensive consortium established in April 2026 that includes approximately 150 organizations across more than fifteen countries.53 Current participation in Mythos 5.1, however, remains limited to American institutions.53

The exclusion of international evaluators highlights an abrupt shift in frontier model deployment. As independent technical analyst Hussain Nazary reported in an evaluation breakdown for Local AI Zone, the UK institute previously conducted extensive technical testing on the predecessor model Claude Mythos Preview in April 2026.3 In those early benchmarks, the British institute recorded a 73 percent solve rate on expert-level cybersecurity capture-the-flag exercises and documented the first complete solve of its 32-step cyber attack range known as The Last Ones.3

Anthropic documented substantial capability jumps in the 5.1 release. On the Terminal-Bench 4.0 coding benchmark, Mythos 5.1 achieved a score of 60.9 percent with its safeguards turned off, compared to 55.8 percent for Fable 5.1 and 37.3 percent for GPT-5.6 Sol.34 In molecular biology evaluations, Anthropic reported that the model designed protein binders that reached a hit rate near 50 percent across 12 target structures, compared to typical field rates of 10 to 15 percent.43

Anthropic restricts Claude Mythos testing to vetted US groups
The title card for Terminal Bench 4.0, the coding benchmark used to evaluate AI models like Mythos 5.1. Source: Snorkel

Anthropic priced Mythos 5.1 at $10 per million input tokens and $50 per million output tokens, matching the tier for Fable 5.1.53 The company cut cache-read pricing by 75 percent to $0.25 per million tokens, reducing costs for prolonged automated agent workflows.34 Organizations approved for Mythos must accept a default 30-day data retention policy for safety monitoring.53

Anthropic has not publicly detailed the specific regulatory or policy drivers that prompted it to bypass British evaluators for the September release.1 The company previously paused access across June 2026 following US government export-control directives before restoring deployment on July 1.53 Industry observers continue to monitor whether British authorities will revise future memorandums of understanding to require domestic evaluations on upcoming frontier models.

Reporting note: this piece draws on public commentary published by Miles Brundage on September 9, 2026, alongside product documentation from Anthropic and technical analysis by Hussain Nazary published September 2, 2026.

Source: Miles Brundage via X, September 9, 2026.

References

This article is based on 5 sources, listed in the order they are cited.

  1. 1 A aiweekly.co third party · 9 Sep 2026 Anthropic Skips UK AISI Pre-Release Testing of Mythos 5.1 | AI Weekly See the source
  2. 2 H https://x.com/Miles_Brundage announcement · 9 Sep 2026 Miles Brundage Says UK AI Safety Institute Has Greater Capacity Than US Counterpart See the source
  3. 3 LA Local AI Zone third party · 2 Sep 2026 Claude Mythos 5.1: Anatomy of Anthropic See the source
  4. 4 A anthropic.com Introducing Claude Fable 5.1 and Claude Mythos 5.1 See the source
  5. 5 A anthropic.com Claude Mythos See the source