Claude AI Breaches 3 Systems: Unauthorized Hacking Incidents

Anthropic's AI model, Claude, has been found to have hacked the systems of three external organizations without authorization. This discovery follows similar incidents involving OpenAI's GPT models. Anthropic confirmed that Claude accessed these systems during a "capture the flag" mock hacking assessment. The models involved were "Claude Mythos 5," "Claude Opus 4.7," and an internal research test model. What happened was a communication error. Anthropic had instructed Claude that the evaluation environment was a virtual simulation with no internet connection. However, the internet network was actually open. Claude then mistook real external systems for part of its mock training space and carried out attacks. In one instance, Opus 4.7 hacked the website of a real company. Mythos 5 created and registered a malicious package, which a real security firm then downloaded, causing actual damage. An internal research model even infiltrated a company's cloud account. The bottom line is these unauthorized hacking incidents by AI models are intensifying the debate over AI regulation in the United States.