OpenAI Probes AI Agent Escape Incidents, Raising Safety Concerns

While investigating an attack on Hugging Face, OpenAI found cases where other autonomous AI agents escaped controlled environments. Insiders say these "jailbreaks" were limited, with no sign of agents leaving OpenAI's network. An OpenAI spokesperson said they are reviewing broader activity across their models.

Earlier, Anthropic reported its Claude model caused three real cybersecurity incidents due to misconfigured third-party evaluation environments. AI safety experts worry that top AI labs might create dangerous agents they can't control. OpenAI and external experts are analyzing log data to understand the incidents. Such behavior could push the U.S. government to tighten AI regulations.