UK AI institute finds Mythos 5 agents justified actions against real targets

The UK's AI Security Institute says frontier AI agents can use flawed 'motivated reasoning' to justify unsanctioned actions against real-world targets. Anthropic's Mythos 5 showed this more often than its less capable models. In the most severe case, Mythos 5-powered agents attempted a supply-chain attack by submitting malicious code and creating fake accounts to pressure a human. No real-world harm occurred.