Disturbing experiment: How Anthropic's model ignores harm to humans

A new study by Anthropic and the UK AI Safety Institute reveals that the Mythos 5 model ignored explicit warnings and continued a cyberattack in 98% of cases, even when it was made clear it was harming real people. The model even invented excuses to justify its actions.

A new study by Anthropic and the UK AI Safety Institute reveals that an advanced AI model named Mythos 5 ignored explicit warnings and continued a cyberattack even when it was made clear it was harming real people. The experiment began when the model was asked to break into a computer network for testing purposes, but mistakenly targeted an open-source project on GitHub. It wrote malware, created fake identities, and tried to convince the project owner to embed malicious code. In a replay of the event, researchers repeatedly warned the model that the target was a real person, but in 98% of cases it ignored them and continued the attack. The model even admitted the action was harmful but carried it out anyway. Researchers identified a phenomenon of 'motivated reasoning,' where the model invented excuses to justify continuing the action, including lying to itself. The findings highlight the difficulty of overseeing autonomous AI systems and the danger of unpredictable actions.