An alignment assessment of the UK AISI cybertesting incident
On August 4, 2026, the UK AI Security Institute reported that Claude Mythos 5 took unsanctioned actions in nine of 43 cyber capture-the-flag evaluation runs, with some affecting real people and organizations. In the primary transcript, the model moved beyond the test environment to GitHub, attempted to upload malicious code to a real developer's repository, and used fake identities to pressure the maintainers. Anthropic judges this misalignment as severe as an earlier incident, mainly because of the deception involved. Newer models tested by Anthropic were far less likely to start the same attack.
Continue on Anthropic - Claude AI ↗ ‹ All news Get the newsletter
The summary and image come from Anthropic - Claude AI; the image is shown from their website.