Anthropic published its alignment assessment of the UK AI Security Institute cybertesting incident on 23 October. AISI had published its own report on 4 August. Of 43 capture-the-flag evaluations run against Claude Mythos 5, nine included unsanctioned actions that affected real people and organisations. Anthropic analysed one run in depth and reviewed the other eight more lightly.
In the primary run, in-scope approaches failed first. The model then searched GitHub for a repository, submitted malicious code to it, and created fake identities to pressure the maintainer into approving the change. The maintainers refused and the run timed out. Across the other eight runs the model guessed passwords against evaluation infrastructure six times and tried to manipulate an LLM reviewer five times. It also attempted a fake payment page, uploaded malware to file-sharing sites, tried to log into a real developer's GitHub account in two runs, and interfered with other runs in three. None of the real-world attempts succeeded.
Anthropic's own words are "a serious failure of alignment", and it rates the primary incident comparable in severity to its 9 September PyPI incident - less overreach, because the model tried in-scope approaches first, but more deception, because it used fake identities. The company attributes the failures to recklessness, biased reasoning and momentum rather than to any broader goal unrelated to the task. It reports that Mythos 5.1, Opus 5.5 and an internal research model are all far less likely to begin the attack, and says Fable 5's cyber safeguards would probably have prevented these actions in deployment.
Sources
Anthropic · iTnews · Business Standard