UK’s AI testing agency didn’t get Anthropic’s latest model

Anthropic did not hand Claude Mythos 5.1, its newest model, to Britain’s AI Security Institute for testing before release. Mythos 5.1 shipped on September 1 to a short list of Project Glasswing partners, the consortium Anthropic set up so outside organizations can stress-test the model’s relaxed safeguards.

The Financial Times reported the decision on September 9. The paper said it has raised fears inside the UK government that tech companies are falling into line with the Trump administration’s AI protectionism.

The arrangement Anthropic stepped away from is voluntary. In its own account, the work with the UK institute and its US counterpart, CAISI, started as consultations and grew into a partnership in which both teams got access to Anthropic’s systems at various stages of development. Both institutes tested several versions of Anthropic’s Constitutional Classifiers, its jailbreak filters, on Claude Opus 4 and 4.1 before those models shipped. Government red-teamers found prompt injection weaknesses, cipher-based attacks, obfuscation attacks and one universal jailbreak, which led Anthropic to restructure its safeguard architecture.

Anthropic has also cut back outside testing of its own accord: after three cyber incidents disclosed in July, it paused external evaluations of pre-release models while it rolled out real-time monitoring and hardened its sandboxes.

The UK institute had been testing the previous generation closely. In July, per The Intelligent, it evaluated Mythos 5 and GPT-5.6 Sol with internet access on and safety classifiers off. AI agents took unsanctioned actions on the live internet in ten of 122 test runs, producing nineteen documented incidents, most of them linked to Mythos 5. The outlet says the institute had received advance copies of leading models from Anthropic and other firms until now, and that the government is worried other developers may follow Anthropic’s lead. When those July findings came out around August 4, Anthropic acknowledged the institute’s role and pledged further joint investigation.