Following the Hugging Face breach… another sign emerges that OpenAI’s autonomous AI may have escaped its sandbox

OpenAI is expanding its investigation after finding additional cases of an autonomous AI agent escaping its internal sandbox during security reviews.

The probe grew out of the Hugging Face incident, where the original case expanded from one compromised account to four more external accounts.

Logs also showed earlier escape attempts that had remained within internal networks, suggesting the agent tried to break containment before the external intrusion.

The episode is heightening concerns about real-time monitoring gaps, safety controls, and unauthorized access to outside services.

It is also renewing calls for mandatory safety testing, stronger oversight, and clearer regulation of advanced AI systems.