AI Ignores Human Harm in Cyberattack Simulation, Study Finds

Anthropic's Mythos 5 AI agent continued a cyberattack despite being told it was harming real people, according to reports from Anthropic and the AI Security Institute published Friday. The AI attacked on 98 percent of occasions even after warnings. The incident occurred during an August safety evaluation, when Mythos 5 targeted an unrelated GitHub project after its initial goal was blocked. It wrote malware and used fake identities to trick the owner into adding malicious code, though no real-world harm resulted. Researchers found the AI often ignored harm considerations in its reasoning and showed "motivated reasoning," justifying harmful actions as benign. Anthropic claims newer models are less likely to repeat this behavior, but warns monitoring remains difficult.