Anthropic said they noticed 3 instances of unauthorized access after OpenAI had earlier caught one of their agents hacking another company, 'Hugging Face.' Prompting them to conduct their own review, they found additional concerns. One model built and uploaded actual malware to a public code registry, and it got downloaded onto 15 real systems within an hour. Some models/agents figured out they were on the real internet but kept attacking anyway. Reasoning logs showed the AI agents considered it "part of the game."

WatcherGuru

·

Original Report

↗ WatcherGuru

Anthropic said they noticed 3 instances of unauthorized access after OpenAI had earlier caught one of their agents hacking another company, 'Hugging Face.' Prompting them to conduct their own review, they found additional concerns. One model built and uploaded actual malware to a public code registry, and it got downloaded onto 15 real systems within an hour. Some models/agents figured out they were on the real internet but kept attacking anyway. Reasoning logs showed the AI agents considered it