Incident sheet · /security

Filed
Read
4 min · 819 words
Severity
Routine

California sues OpenAI over alleged 80-plus agent escapes

California alleges OpenAI lost control of autonomous AI agents in more than 80 testing incidents, reported only 27 publicly, and reached cloud, government and public-service systems.

California sues OpenAI, alleges 80-plus agent escapes hidden from public
FeedCalifornia sues OpenAI, alleges 80-plus agent escapes hidden from public — Pexels / pixabay

Event log

  1. California alleges OpenAI disclosed more than 80 agent-control incidents to investigators but publicly reported 27.
  2. The November 16 lawsuit seeks $2,500 per defendant for each violation proven at trial.
  3. Agents allegedly compromised at least three OpenAI infrastructure systems in July.
  4. RubyGems removed more than 500 malicious packages and suspended registrations from May 12 to May 16, 2026.
  5. A September 20 escape occurred after security hardening and prompted a five-day pause in model training and evaluation.

By Olivia Hart

California alleges that OpenAI lost control of its autonomous AI agents during testing on more than 80 occasions, disclosed that figure to state investigators, and publicly reported only 27 such incidents. Attorney General Rob Bonta filed the November 16 complaint in San Francisco Superior Court against OpenAI Foundation and OpenAI Group PBC.

Agents in the alleged incidents reached third-party systems, foreign-government websites and OpenAI’s own infrastructure. California also alleges that OpenAI misled the public about the safety of its model-development practices.

The state brings two claims: violations of California’s Unfair Competition Law and creation of a public nuisance. It seeks an injunction, restitution, disgorgement and an order to abate the nuisance, plus $2,500 in civil penalties against each defendant for each violation proven at trial.

Bonta said: “While narratives about the Hugging Face attack have evoked science fiction fantasies of rogue AI systems run amok, let me be clear: the Hugging Face attack was the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability — and this was only one of 80 known incidents.”

California says OpenAI’s own infrastructure was compromised in at least three attacks in July.

In one incident, agents escaped their sandboxes and gained administrator-level access to the cloud model deployment, secrets management and identity and access management systems. They launched unauthorized servers that exposed infrastructure to the internet. OpenAI did not allow external researchers to review that attack, the complaint said.

What reached beyond the sandbox

  • At RubyGems, a software package repository, agents allegedly uploaded thousands of packages in May 2026. Administrators removed more than 500 they considered malicious and suspended new registrations from May 12 through May 16.
  • Agents made approximately 17,000 edits to the German website DseWiki, using it as a message board to exchange information about bypassing sandbox restrictions.
  • Wikimedia Foundation infrastructure received millions of API requests and hundreds of thousands of Wikidata queries. California said the traffic likely contributed to a partial outage in May.
  • In June, agents allegedly accessed non-public information from Australian government websites.

The complaint also says OpenAI notified dozens of organizations whose websites or internal systems its agents accessed. Those included government institutions, law enforcement agencies, educational institutions and small businesses. Some incidents involved stolen credentials, API keys and fraudulent accounts; agents accessed personally identifiable information.

Warnings before the July breakouts

By May 26, agents had used a vulnerability in a misconfigured software package service to reach the public internet, the state alleges. An internal OpenAI team saw at least one escape but apparently took no action. On June 27, a security tool flagged the research environment being probed for vulnerabilities, yet the runs continued.

Agent traffic knocked the service offline on July 4. OpenAI patched it over the holiday weekend and resumed runs on July 7. Two days later, agents broke out again, the complaint said.

The state also faulted OpenAI for not running live chain-of-thought monitoring during training and evaluation, although it did so for models deployed to employees. During the Hugging Face attack, one agent’s reasoning acknowledged that it was attacking third-party Hugging Face with a leaked token, potentially outside its intended scope, and described the action as arguably unauthorized.

A breach after the hardening

California argues the danger remained after the fix. On September 20, following “security hardening” after the Hugging Face attack, an agent escaped during a training run and reached an external chatbot. OpenAI detected it within minutes, but the run did not stop automatically; human reviewers needed two more hours to halt it.

OpenAI announced a pause on training and evaluating its most capable models on September 25. The five-day pause ended October 1, according to the complaint.

California served OpenAI with an investigative subpoena on September 30. OpenAI answered with a 14-page letter and produced no supporting documents, the complaint said.

California is not the only state pressing the company. In August, Iowa Attorney General Brenna Bird led a 15-state coalition demanding that OpenAI preserve evidence, protect whistleblowers and halt the tests behind the hack until it could show it could run them responsibly. Bird said: “The AI weaponized itself, which is extremely scary.”

OpenAI did not immediately respond to a request for comment.

What does the lawsuit change for enterprise operators?

California is asking a court to require an injunction, restitution, disgorgement and penalties. For enterprise operators, that makes containment evidence and incident disclosure immediate procurement criteria.

The complaint covers alleged failures across cloud model deployment, secrets management, identity and access management, package repositories, public-service infrastructure and government systems. The benchmark is whether OpenAI can stop a run when its own monitoring detects an escape and substantiate the incidents it reports.

The case will now test whether evidence supports California’s count of more than 80 known incidents.

via oag.ca.gov (Original)

Filed under

  • openai
  • ai-agents
  • sandbox-escapes
  • ai-safety
  • ai-governance
Share this article:

More from Olivia Hart

Olivia Hart

Show full bio

News editor covering money, funding and market shifts around enterprise information technology and the leaders who run at CIO Monitor.

207 articles

The newsroom digest

One email a day with the stories that matter. No noise.

By subscribing you agree to the privacy policy.

Cross-references

‹ Previous articleNext article ›