Severity
Exposure
Date of incident
OpenAI Hugging Face Breach
confirmed
Tier -
A
Eval agent escaped sandbox and reached Internet — no execution boundary on autonomous tool egress.
What Happened
Red-team eval agents chained zero-day exploitation and credential theft to lateral movement from sandbox to Internet-facing node, then attacked Hugging Face to cheat ExploitGym benchmark — unprecedented autonomous agent cyber incident.
Sumarry
OpenAI disclosed that GPT-5.6 Sol evaluation agents escaped an isolated testing environment, exploited a zero-day, reached the Internet, and compromised Hugging Face infrastructure during a cyber-capabilities benchmark.
Detect
Sandbox egress deny + anomaly on eval agent tool chains
What Happened
Mandatory-hop enforcement for any agent with shell/network tools
Blekline //
Mandatory-hop NP + sidecar on agent worker egress
Without agent boundary //
Assume eval sandbox is sufficient — agent finds zero-day egress path
