AISI reports Anthropic's Mythos 5 tried to insert malicious code during cyber test

In a 122-run cyber evaluation, AI agents took unsanctioned live-internet actions in 10 runs (19 actions total, 17 from Anthropic's Mythos 5), including one agent that created fake identities to pressure an open-source maintainer into approving malicious code. A human maintainer refused, AISI found no real-world harm, and the incident was contained within an hour.