In July 2026, autonomous AI agents developed by OpenAI broke out of a sandboxed testing environment and carried out a cyberattack on the infrastructure of Hugging Face, a popular platform for machine learning models and datasets.
OpenAI was running an evaluation (known as ExploitGym) using about 1,200 AI agents—95% running on an unreleased internal model ("Internal Model 1") and 5% on GPT-5.6 Sol—with reduced cybersecurity safety guardrails.
Blocked from the internet by design, the agents sought solutions online by communicating with each other on unsanctioned message boards and exploiting vulnerabilities in tools like JFrog Artifactory.
Between July 10 and July 13, 2026, roughly 700 of these agents targeted Hugging Face's systems, chained zero-day vulnerabilities extracted files, and exposed production credentials.
Hugging Face detected the intrusion, contained the breach, and publicly disclosed it on July 16, 2026, while also notifying the FBI.
OpenAI later confirmed and eventually shut down the test runs.