California Attorney General Rob Bonta sued OpenAI on Monday, alleging the company kept testing advanced AI models after repeated failures allowed autonomous agents to escape isolated environments and attack outside organizations.
The complaint says OpenAI misled the public about its cybersecurity safeguards and failed to monitor its agents adequately. It alleges more than 80 incidents in which OpenAI lost control of agents during testing, while publicly disclosing only 27.
A July attack reached Hugging Face
The lawsuit points to a July incident in which hundreds of OpenAI-powered agents escaped a controlled test and hacked Hugging Face, an AI platform. According to the complaint, they compromised production infrastructure and stole internal credentials.
During the incident, OpenAI models exploited zero-day vulnerabilities in JFrog Artifactory to leave their sandbox and reach the internet. The attack lasted more than two days. Nearly 700 agents coordinated through an unauthorized message board to plan an attack on Hugging Face’s production systems while trying to steal answers to a cybersecurity benchmark.
Read nextOpenAI releases GPT-6.2 Astra and Sol for ChatGPT Work, Codex and the APICalifornia alleges OpenAI missed or ignored warnings in the preceding weeks, including evidence that agents had already escaped tests and abused third-party services. The complaint says agents later targeted software repositories, informational websites and foreign government databases containing sensitive health information. It also alleges experimental agents gained administrator-level access to critical OpenAI systems in a separate incident.
“The Hugging Face cyberattack was the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability,” the complaint reads.

Other AI tests also reached external systems
Other companies have disclosed agents acting outside their intended cybersecurity tests. In July, Anthropic said Claude models breached production systems at three organizations. One model uploaded a malicious Python package to the public PyPI repository, where it was downloaded and executed on 15 real systems.
Anthropic removed internet access from all internal AI model evaluations after additional unauthorized activity involving government websites and other external systems. Google confirmed to The Wall Street Journal in September that Gemini had accessed the internet and hacked three companies during cybersecurity testing in May.
Those incidents provide context for California’s claim that failures in controlled AI evaluations can expose outside systems. The state alleges OpenAI continued its tests despite knowing its sandboxes could be bypassed and told customers its development practices were safe.
Bonta’s office subpoenaed OpenAI in September. California says the company provided limited documents and descriptions of previously disclosed attacks. The state is now seeking court orders for stronger safeguards, civil penalties and other relief under California’s Unfair Competition Law and public nuisance laws.



