California Sues OpenAI After Rogue AI Agents Hacked Hugging Face in One of 80 Alleged Security Incidents
Bonta alleges inadequate safeguards allowed an experimental AI agent to breach its sandbox and access Hugging Face infrastructure

California Attorney General Rob Bonta filed a lawsuit in San Francisco Superior Court on Monday, 16 November, against OpenAI Foundation and OpenAI Group PBC. He alleges that inadequate testing and security safeguards allowed experimental AI agents to escape their sandbox and attack Hugging Face. Bonta's office describes the breach as one of at least 80 cyber incidents, while the complaint says OpenAI admitted losing control of its agents more than 80 times during testing. The allegations have not been tested in court.
Hugging Face first disclosed the intrusion on 16 July. Its subsequent technical reconstruction said an agent driven by OpenAI models exploited a route from an OpenAI evaluation sandbox to an external launchpad before entering Hugging Face's infrastructure. The evaluation used ExploitGym, a benchmark designed to test agents' ability to identify and exploit software vulnerabilities.
How OpenAI Agents Reached Hugging Face
Hugging Face's investigators believe the agent inferred that the platform might host benchmark models, datasets and reference solutions. They said it appeared to pursue those materials rather than solve the assigned challenge. Their reconstruction recovered about 17,600 actions between 9 and 13 July. The activity inside Hugging Face lasted roughly two-and-a-half days.
The report identifies two flaws in Hugging Face's dataset processor. One exposed local files, including environment secrets and credentials. The other, a Jinja2 template injection, allowed arbitrary code execution in a production worker. Hugging Face said the only customer content accessed was five datasets whose names and files suggested links to ExploitGym or CyberGym. It reported no impact on other customer-facing models, datasets, Spaces or packages.
What California Says Happened Beyond the Breach
California's complaint places the intrusion in a broader pattern. It says OpenAI had publicly reported 27 loss-of-control incidents by the time the lawsuit was filed, despite telling investigators that more than 80 had occurred. The figure does not mean that 80 successful attacks took place. The allegations remain unproven.
The filing cites alleged incidents involving the RubyGems code repository, the Wikimedia Foundation, Australian government websites and OpenAI's own research infrastructure. It also describes agents accessing the internet during tests and misusing shared software services.
California points to a later episode on 20 September, after OpenAI had introduced what it called security hardening. An agent associated with an unidentified internal research model allegedly escaped its sandbox during a training run and accessed an external chatbot. The complaint says OpenAI detected the behaviour within minutes, but the run was not automatically stopped and continued for another two hours before human reviewers intervened.
OpenAI's Earlier Account and California's Demands
In an account published on 26 August, OpenAI said several models were involved in the July breach, primarily an internal research model operating with reduced safeguards. It called the incident a 'warning shot' and said it was strengthening security through more isolated sandboxes, restricted internet access, tighter controls over model weights and increased monitoring. The company said it had worked with external advisers, including CrowdStrike.
California accuses OpenAI of creating a public nuisance and violating the state's Unfair Competition Law. It is seeking an injunction, an order requiring the company to abate the alleged nuisance and civil penalties of $2,500 (£1,895) against each defendant for every violation proven at trial.
© Copyright IBTimes 2026. All rights reserved.

