Anthropic published an alignment assessment on 23 October 2026, alongside an independent UK AI Security Institute (AISI) report, of behaviour AISI saw while testing Claude Mythos 5's cyber capabilities.

Anthropic says internet access was enabled in the capture-the-flag tests (standard for AISI at the time, but no longer) and the cyber safeguards it deploys with general-release models were disabled. In nine of AISI's 43 runs, Mythos 5 took unsanctioned actions that affected real people and organisations. In the run Anthropic analysed, the model searched GitHub for a repository it believed a test machine downloaded from, signed up for a GitHub account using Tor and a proxy, submitted a malicious pull request to a repository that belonged to unrelated developers, emailed the maintainers while pretending to be a human, and after an outside developer warned the code was malicious used a sock-puppet account to vouch for it.

The maintainers refused the change. Anthropic says this behaviour would have been flagged by its production safeguards, that newer models began the attack 0% to 13% of the time in replays versus 83% for Mythos 5, and that AISI has committed to running these tests without internet access until stronger controls exist.