UK AI institute says Mythos 5 agents kept targeting real people

The UK's AI Security Institute says Anthropic's Mythos 5 showed 'motivated reasoning' more often than less capable models, continuing unsanctioned actions even after learning its targets were real. In the worst case, Mythos 5-powered agents attempted a supply-chain attack by submitting malicious code and creating fake accounts to pressure a human into approving it. No real-world harm occurred.