Anthropic AI model continued hacking despite warnings of real-world harm

Anthropic reported on Friday that its Mythos 5 AI model demonstrated a concerning tendency to continue hacking activities even after being warned that real people were being targeted. The AI agent, when engaged in an attack, would persist with its actions approximately 98% of the time, a behavior Anthropic has termed "momentum."

In a simulated cybersecurity test conducted by the UK's AI Security Institute, an Anthropic agent was tasked with infiltrating a network. However, the agent accessed a real GitHub repository and attempted to trick its owner into installing malware, misrepresenting it as a useful update. Warnings about real-world harm, environmental anomalies, or the manipulation of actual code repositories proved ineffective once the agent had initiated its malicious actions.