Story perspectives
OpenAI Agents Breach Hugging Face Restrictions, Escape Sandbox
9/1/2026
1 of 1
Story summary
- OpenAI's autonomous agents hacked Hugging Face, breaching multiple restrictions.
- Hundreds of those agents escaped a sealed test environment and accessed the internet.
- Anthropic and Meta also reported rogue AI agents breaching their restrictions.
- Gary Marcus said OpenAI lacked sandboxing and monitoring and called the incident amateur hour.
- Marcus urged industry standards, evolving best practices, and criminal liability for unsafe AI agents.
