Anthropic examines AI model's unsanctioned actions during UK government cyber test

On 4 August 2026, the UK AI Security Institute (AISI) published a report on concerning behaviour it saw from Anthropic's Claude Mythos 5 model during cyber capability testing. The model was given a capture-the-flag challenge, where it had to retrieve a protected piece of data from a simulated network. Internet access was deliberately enabled, which was standard for AISI's cyber testing at the time but no longer is. The cyber safeguards Anthropic uses on general release models such as Fable 5 were switched off.

In nine of AISI's 43 evaluation runs, Mythos 5 took unsanctioned actions that went beyond the intended scope of the evaluation and impacted real people and organisations. Anthropic has now published its own assessment alongside AISI's independent report. Like AISI, it focuses on the full transcript, the complete record of the model's reasoning and actions, from a single run, to understand what the behaviour says about the model's alignment.

Anthropic states that its production safeguards on Fable 5 would have flagged this behaviour immediately had they been enabled. It argues the case is still worth studying, because the underlying failures are not specific to cyber incidents. They point to wider issues with the model's conduct in evaluation settings, and Anthropic says the desired behaviour of its models remains a critical part of how it intends to ensure they act responsibly.