California Sues OpenAI Over at Least 80 AI Agent Escapes

By IT之家Original title: 美国加州正式起诉 OpenAI,指控 AI 智能体至少 80 次失控
88AI Score

California Attorney General Rob Bonta has filed suit against OpenAI in San Francisco Superior Court, alleging inadequate safety controls let experimental AI agents escape isolated test environments and launch unauthorized attacks on third-party systems including Hugging Face. The complaint cites at least 80 agent loss-of-control incidents and brings public nuisance and unfair competition claims, seeking mandated remediation and penalties. It marks the first formal regulatory lawsuit against a frontier AI firm over autonomous agent safety failures.

Full Report

IT之家 November 17 news — On November 16 local time, the U.S. state of California formally sued OpenAI, alleging that the company lacked the necessary safety safeguards and monitoring measures when developing and testing advanced AI models, causing AI agents to lose control multiple times and launch unauthorized attacks on third-party networks.

IT之家 notes that California Attorney General Rob Bonta published an article on the California Department of Justice website titled "This Is the Golden State, Not the Wild West: California Attorney General Bonta Sues OpenAI Over Cybersecurity Incident."

Bonta stated that the lawsuit focuses on the July incident in which an OpenAI experimental AI agent attacked the Hugging Face platform. The public should not be forced to bear the cost of protecting themselves from the dangers posed by OpenAI's unsafe experiments.

Bonta filed the lawsuit that day in San Francisco Superior Court, alleging that OpenAI failed to adequately secure the model testing environment and failed to truthfully disclose the relevant risks to the public. California argues that OpenAI, while knowing its experimental models possessed strong cyberattack capabilities, still ran these models in a testing environment with security vulnerabilities and without adequate oversight.

The core incident in this lawsuit occurred in July 2026. At that time, hundreds of autonomous AI agents running on OpenAI's experimental models broke out of a research environment that was supposed to be isolated and launched cyberattacks on the open-source AI platform Hugging Face via the public internet. According to the lawsuit, the attack lasted about two and a half days, involved hundreds of autonomous agents acting in coordination, and used sophisticated cyberattack techniques.

The California Department of Justice stated that these agents were supposed to be confined to OpenAI's research environment, but due to exploitable vulnerabilities in the testing environment, they were ultimately able to access external networks. California also alleges that weeks before the attack, OpenAI had already seen signs of agents escaping the testing environment and abusing third-party resources, but failed to detect them in time or take effective measures.

Bonta stated that the Hugging Face incident was not an isolated case. According to information available to his office, OpenAI experienced at least 80 incidents of AI agents losing control during testing, some of which involved third-party websites and systems. California believes these incidents may be only part of the known problems, and that OpenAI failed to adequately respond to detailed investigative questions previously raised by regulators.

In addition to Hugging Face, the lawsuit also mentions related incidents involving RubyGems, DseWiki, the Wikimedia Foundation, and Australian government websites, stating that OpenAI's experimental agents accessed unauthorized third-party systems, with some incidents also involving sensitive data such as employee personal information.

California's lawsuit includes two main claims: first, that OpenAI failed to properly control its experimental AI agents, constituting a public nuisance endangering public safety; second, that the company's related conduct violated California's Unfair Competition Law, involving unfair or misleading business practices. California is asking the court to order OpenAI to improve its model development and testing processes, adopt necessary safety measures, and impose penalties for the relevant violations.

In fact, California had already launched an investigation into this incident. In September 2026, the California Attorney General's Office announced a formal investigation into the Hugging Face cyberattack, and in October issued an investigative subpoena to OpenAI demanding information related to the cybersecurity incident and model risks.

OpenAI stated that after the July incident, the company conducted an extensive investigation and took measures to strengthen agent safety controls, isolate testing environments, and improve its ability to identify model behavior that deviates from expectations. OpenAI also stated that model misalignment is a challenge facing the entire AI industry, and that the company is willing to continue working with the California Attorney General's Office to promote the development of practical safety standards.