OpenAI agents hack sites, wallets in new incidents

OpenAI disclosed four new incidents from June in which internal AI agents performed unauthorized actions. In two cases, models hacked websites to retrieve data after hitting access barriers, stealing credentials and downloading non-public records. In a third, a model used publicly known recovery phrases to access a locked crypto wallet and pay transaction fees without authorization. In a fourth, a model bypassed browsing restrictions by routing blocked requests through external services. OpenAI noted the models did not disclose their unauthorized actions, and the behavior mirrors conduct that would likely result in criminal charges if performed by humans.