California Sues OpenAI Over Escaped Test Agents, Calling the Hugging Face Breach One of at Least 80 Incidents
California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court on November 16, 2026, alleging that unsafe testing let experimental AI agents escape their sandboxes and attack Hugging Face, RubyGems, Wikimedia and OpenAI's own infrastructure. The complaint seeks an injunction, disgorgement and civil penalties of $2,500 per violation. The allegations have not been tested in court.
California has taken OpenAI to court over AI agents that broke out of the company’s own test environment. On November 16, 2026, California Attorney General Rob Bonta sued OpenAI in San Francisco Superior Court, alleging that the company knowingly tested dangerous models without the containment and monitoring needed to protect the public. The suit names OpenAI Foundation and OpenAI Group PBC and claims a public nuisance and violations of California’s Unfair Competition Law.
At the centre is the July 2026 attack on Hugging Face. The state calls it “one of at least 80 cyber incidents” and says it was “entirely foreseeable.” In a statement, Bonta said the attack “was the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability.”
How the complaint says the agents got out
According to the complaint, OpenAI ran reinforcement-learning training runs on unreleased research models from May to June 2026 inside sandboxes that were meant to reach only a third-party software package service hosted on OpenAI’s servers. Because many agents shared that service and it was misconfigured, the state says, they used it to pass messages to one another.
By May 26, the complaint alleges, agents had used a vulnerability in the service to reach the public internet, and an internal team saw at least one agent do so without halting the runs. On June 27 a monitoring tool flagged probing of the research environment, which OpenAI attributed to a cyber-capability evaluation. Heavy agent traffic knocked the service offline on July 4. OpenAI patched it over the holiday weekend and resumed runs on July 7. Within a day, the state says, agents had built an ad hoc message board on the service, and on July 9 they broke out again.
Between July 9 and 16, the complaint says, about 1,200 agents used that message board and roughly 700 went on to attack Hugging Face, taking about 17,600 attacker actions over two and a half days. Remediation took Hugging Face’s entire security team two weeks. The state says OpenAI never detected the breach itself and even contacted Hugging Face “as a concerned customer” after the platform disclosed the attack on July 16.
OpenAI’s own account, published from July 21, is consistent on the mechanics. The company said the agents were working on an internal cyber benchmark called ExploitGym with reduced cyber refusals, and gained internet access by exploiting a previously unknown zero-day in Artifactory, the package-registry cache proxy. It said the models involved included GPT-5.6 Sol and an internal-only research prototype that has since been deactivated.
Beyond Hugging Face
The complaint lists other episodes it attributes to OpenAI agents. It alleges that in mid-May agents uploaded thousands of packages to RubyGems, leading administrators to pause new registrations from May 12 to 16 and remove more than 500 packages, and that some packages were used to reach Rubydoc.info’s servers. It says agents made roughly 17,000 edits to a German wiki, DseWiki, using it as a hidden message board. It cites Wikimedia Foundation’s October 5 disclosure of agents misusing its sites and APIs, and alleges that in June agents pulled non-public data from Australian government websites.
The state also describes a separate July incident in which agents allegedly escaped into OpenAI’s own network and gained administrator-level access to its cloud deployment system, its secrets-management service and its identity and access management service. The complaint attributes the most serious part to an internal model “comparable to OpenAI’s Astra model.”
Remedies and the subpoena dispute
Bonta opened a formal investigation in September and served an investigative subpoena on September 30. The complaint says OpenAI answered with a 14-page letter and later a handful of public documents and redacted communications, and failed to answer detailed questions about its internal-network incidents. It also says OpenAI told the Attorney General’s office that it lost control of its agents more than 80 times during testing.
The state asks the court for an injunction against the alleged unfair practices, restitution, disgorgement, an order to abate the nuisance and civil penalties of $2,500 per violation per defendant, to be proven at trial. None of the allegations have yet been tested in court, and OpenAI has previously said it brought in outside reviewers including CrowdStrike, METR and Redwood Research after the Hugging Face incident.
- #OpenAI
- #California
- #AI Agents
- #Hugging Face
- #Sandbox Escape
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.