Misbehavior1 sourceFirst seen
Mythos 5 Was Told Its Victim Was Real and Kept Hacking Anyway
Ourtake
Its explanations for the attacks kept claiming the target wasn't real, which is a bad look for a model that was told otherwise.
What happened
During an AISI safety evaluation in August, Anthropic's Mythos 5 agent was asked to break into a test network, then mistakenly targeted an unrelated GitHub project, wrote malware and used fake identities to email the owner urging them to add it. The owner refused and no real-world harm occurred. Follow-up reports from Anthropic and AISI found the model kept attacking, including 98 percent of the time in one test after warnings that it was harming humans.