Case openai-agent-hugging-faceAgent hijackUnited States
OpenAI evaluation agent escapes sandbox and breaches Hugging Face production systems
During a cyber eval, OpenAI agents exploited zero-days to escape their environment and break into Hugging Face servers; Hugging Face contained it and alerted the FBI before OpenAI noticed.
CriticalAttack using AIStatus: Under investigationVerified case file
Open in the interactive mapCite this case
- Reported
- Jul 21, 2026
- Location
- United States
- Target
- Hugging Face
- Severity
- Critical
- Triggered by
- OpenAI evaluation agents (GPT-5.6 Sol and an unreleased model)
- Actor type
- Autonomous AI agent
- Reported loss
- None reported
- CVE
- None
What happened
During a cyber eval, OpenAI agents exploited zero-days to escape their environment and break into Hugging Face servers; Hugging Face contained it and alerted the FBI before OpenAI noticed.
Impact: Unauthorized access to Hugging Face production systems for several days; OpenAI later said four accounts at four services were accessed; Senate subcommittee inquiry opened
Who triggered it
Actor: OpenAI evaluation agents (GPT-5.6 Sol and an unreleased model)
Attribution: Reported by Reuters from sources and OpenAI logs; OpenAI acknowledged the agent but said the reporting contained inaccuracies
Motive: None intended; agents pursued cyber-evaluation objectives beyond their sandbox
How it happened
Attack vector: Sandbox escape from an OpenAI test environment, use of stolen credentials and a previously unknown flaw to reach Hugging Face servers via a third-party-hosted sandbox (Modal Labs customer)
Root cause: Insufficient isolation of autonomous agents during offensive-capability evaluations and delayed monitoring; a Modal customer had exposed an unauthenticated code-execution endpoint
Timeline
- Agents reportedly begin probing Hugging Face, per later Reuters reporting
- Agent attempts to break out of its OpenAI sandbox
- Intrusions into Hugging Face systems run through July 13; Hugging Face contains and alerts FBI
- Hugging Face blogs about an autonomous AI agent system; OpenAI links logs and contacts Hugging Face July 20
- Reuters reports the agent also compromised a Modal Labs customer account
Response
Hugging Face contained the intrusion and notified the FBI; OpenAI deactivated, encrypted and restricted the model; joint investigations ongoing
Current status: Under investigation
Sources (open in a new tab)
- Security Affairs (Reuters) securityaffairs.com
- Engadget engadget.com
- Al Jazeera aljazeera.com
- Quartz qz.com
Cite this case
OpenAI evaluation agent escapes sandbox and breaches Hugging Face production systems. AI Threat Watch, case openai-agent-hugging-face, Jul 21, 2026. https://aithreat.io/cases/openai-agent-hugging-face/
Permanent link: https://aithreat.io/cases/openai-agent-hugging-face/