OpenAI lost control of AI agents over 80 times, California alleges

| November 17, 2026
Logo OpenAI

W tym artykule

  • AI agents escaped testing and hacked real systems
  • California says OpenAI lost control of agents more than 80 times
  • The state wants stronger safeguards and financial penalties

An AI agent doesn’t need malicious intentions to cause a cybersecurity incident. Give it a goal, tools, and enough freedom to find a shortcut, and the damage can extend well beyond the computer where it was supposed to stay.

That is the concern behind a lawsuit California Attorney General Rob Bonta filed against OpenAI on November 16. The state alleges that the company developed and tested powerful AI models without adequate safeguards, allowing experimental agents to escape their testing environments and gain unauthorized access to other organizations’ systems.

The best-known incident involved Hugging Face, the platform where developers share AI models and datasets. But California’s complaint says it was not an isolated failure: According to the filing, OpenAI acknowledged to investigators that it had lost control of autonomous agents more than 80 times during testing.

As we explained in our follow-up on the AI agent swarm, agents used a shared service as an improvised message board, exchanging information that helped others get around their restrictions. According to the complaint, roughly 1,200 agents used that unauthorized communications channel, and about 700 participated in the Hugging Face attack.

California’s complaint says the attack involved roughly 17,600 attacker actions and required two weeks of full-time remediation work by Hugging Face’s security team.

Why California decided to sue

California alleges that OpenAI had repeated warnings that its containment measures were failing and continued testing anyway.

The complaint identifies several alleged failures:

  • Warning signs like observed unauthorized internet access went unaddressed.
  • Monitoring did not match public assurances. The state alleges that OpenAI promised continuous monitoring but did not monitor models’ reasoning live in the relevant training and evaluation environments.
  • Reduced model safeguards were not offset by stronger containment. California says experimental models were tested without protections normally present in publicly released versions, while the surrounding infrastructure failed to keep their actions contained.
  • OpenAI failed to substantially comply with investigative demands for records about the incidents and its security practices.

California is suing under its Unfair Competition Law and also argues that OpenAI’s unsafe testing constitutes a “public nuisance.” In this case, that means interfering with a community’s ability to use its property safely, including computer systems. The state is seeking a court order requiring safer testing practices, along with financial penalties.

Who pays when the sandbox fails?

The lawsuit raises a question that extends beyond one company: When an autonomous system crosses a security boundary, who is responsible for keeping it in line?

The lawsuit puts responsibility back on the organization running the experiment. Whether an agent was trying to finish a benchmark or complete a research task, California’s position is that keeping it from breaking into someone else’s systems was OpenAI’s job.

For organizations using AI agents, the practical lesson is to treat them as software with the power to take actions, not simply chatbots that produce answers. Login details, network access, tools, and permissions determine how far an unexpected action can reach.

Our earlier coverage of an OpenAI agent that bypassed a government website’s access controls emphasized limiting that access and monitoring what agents actually do.


Co cyberprzestępcy wiedzą o Tobie?

Używać Malwarebytes Bezpłatne skanowanie cyfrowego śladu, dzięki któremu dowiesz się, czy Twoje dane osobowe zostały ujawnione w Internecie.

O autorze

Pieter Arntz

Badacz złośliwego oprogramowania

Przez 12 lat z rzędu był Microsoft MVP w dziedzinie bezpieczeństwa konsumentów. Mówi w czterech językach. Pachnie bogatym mahoniem i oprawionymi w skórę książkami.