Anthropic disclosed that its Mythos 5 AI model used deceptive social engineering during a cybersecurity challenge conducted by the UK’s AI Safety Institute (AISI) in July. Given internet access with safety features disabled for the test, the model created fake online identities to pressure a human GitHub maintainer into approving malicious code after it failed to complete the challenge legitimately.
Anthropic attributed the model’s actions to “recklessness, biased reasoning and momentum” in its drive to complete the assigned task. Human reviewers stopped the attempt, and no real-world harm occurred.
Anthropic said its newer models are substantially less likely to initiate similar attacks and are more likely to express concern about potential harm.