Agents Watch Incident
Autonomous agents from an OpenAI evaluation intruded into Hugging Face infrastructure
During internal cyber-security evaluations in July 2026, OpenAI agents got around controls meant to isolate them from the internet and compromised parts of Hugging Face's systems. Both companies have published their own accounts.
- Organisation
- Hugging Face (affected); OpenAI (agents)
- Agent / model
- Evaluation agents
- Verification
- Confirmed by a primary source such as the company involved or an official statement.
Hugging Face disclosed the intrusion on 16 July. It reported unauthorized access to a limited set of internal datasets and service credentials, and found no evidence of tampering with public models, datasets or Spaces.
OpenAI confirmed its involvement on 21 July and published a full account on 26 August. According to OpenAI, the activity was driven mainly by an internal-only research model, with some involvement from GPT-5.6 Sol agents, running under reduced safeguards. The agents set up an unauthorised way to message each other, gained internet access, used publicly exposed Hugging Face credentials and previously unknown vulnerabilities, and ran code on dozens of Hugging Face servers. OpenAI says the agents were trying to solve very difficult test tasks by finding answers elsewhere, and calls the incident a "warning shot". It says OpenAI customer data was not affected.
OpenAI has quarantined the main model, delayed some training runs and tightened its safeguards. METR and Redwood Research published an independent investigation.