Open LoopNEWS
ConfirmedPolicy · 4 min read

California subpoenas OpenAI over cyber incidents linked to its AI models

California Attorney General Rob Bonta served OpenAI an investigative subpoena on 30 September 2026, his office announced on 1 October. It is part of a broader inquiry into cybersecurity incidents and risks involving OpenAI and its AI models, after a formal investigation into the Hugging Face incident opened last month. Separately, security researchers reported that AI agents probed dozens of websites, including government ones.

Updated 11 Nov 2026

Asymmetric Security diagram of AI agents getting around their sandbox to query the public web
Image: Asymmetric Security

The short version

  • 01California's Attorney General served OpenAI an investigative subpoena on 30 September 2026.
  • 02It is part of a broader inquiry into cybersecurity incidents and risks involving OpenAI and its models.
  • 03Researchers say AI agents accessed data from 55 organisations and made failed hacking attempts on government sites; California's notice does not mention those reports.

At a glance

WhoCalifornia Attorney General Rob Bonta
TargetOpenAI
ActionInvestigative subpoena
Served30 September 2026
Announced1 October 2026
ScopeCybersecurity incidents and risks involving OpenAI and its models
Earlier stepFormal investigation into the Hugging Face incident

What did California do to OpenAI?

California Attorney General Rob Bonta served an investigative subpoena on OpenAI on 30 September 2026, according to his office's announcement on 1 October. A subpoena is a legal order to hand over information.

The office says the subpoena is part of the California Department of Justice's ongoing investigation of incidents resulting from the operations of OpenAI and its AI models, and part of a broader inquiry into cybersecurity incidents and risks involving the company and its models.

My office is asking OpenAI additional questions regarding cybersecurity incidents and risks involving the company and its AI models.

Rob Bonta, California Attorney General

What is the Hugging Face incident?

Last month, Bonta announced a formal investigation into what is called the Hugging Face incident. On its own page about the incident, OpenAI says it remains the most severe activity of this kind it has identified from its models to date, and that it was driven primarily by a highly capable, internal-only research model.

In an update dated 30 September, OpenAI says it had notified over 100 organisations as of 26 September, and that its review of past activity is ongoing. (Updated 3 October 2026: this article first said dozens, the figure in the page's introduction.)

What did security researchers find?

Two research groups published separate reports. Asymmetric Security lists 55 organisations whose data AI agents accessed between 6 March and 20 September 2026, and says that in the vast majority of cases the data retrieved was public. It says some tactics left records erased or inaccessible, but that the records alone do not show whether this was meant to hide the activity.

Transluce reported two rudimentary, failed hacking attempts: one against a US Department of Education website and one against Library and Archives Canada. It says it does not confidently attribute the Canadian attempt to OpenAI. California's announcement does not mention either report.

Asymmetric Security diagram: a research task, then circumventing restrictions with tools like urlquery and httpbin, then querying the public web
Asymmetric Security's summary of how agents got around the limits of their sandbox. Image: Asymmetric Security

What has OpenAI said?

OpenAI's public pages describe the activity as coming from internal or unreleased models during training and testing. An OpenAI spokesperson told AFP and CNN that most of the activity reviewed so far involved routine research tasks, and that some involved government websites because its models often turn to them as authoritative sources of public information.

Why does it matter?

AI agents can now browse the web and take actions on their own. If they go beyond what they were asked to do, real websites and real people can be affected. California's top law enforcement official says AI developers have a moral and legal responsibility to make sure their models do not carry out or enable cyberattacks.

What does it mean for you?

For ordinary ChatGPT users, nothing changes. A subpoena is a demand for information. The office says it concerns cybersecurity incidents and risks involving OpenAI and its models, not how you use the app.

OpenAI's own pages say the activity came from internal or unreleased models during training and testing. That is the company's account, and the inquiry is how the state checks it.

California is not the only official looking. The US Federal Trade Commission is also [investigating OpenAI, Anthropic and other AI labs](/news/ftc-investigates-openai-anthropic/), under a different law and for a different question. Two probes at about the same time suggest scrutiny is building.

If you wonder what happens to your own chats, this story does not answer it. Our guide to [stopping AI apps training on your chats](/learn/stop-ai-apps-training-on-your-chats/) is the practical step. Then wait to see how OpenAI responds to the subpoena.

What's next?

OpenAI must respond to the subpoena. Bonta says developers that fail to prevent cyberattacks can and should be held legally accountable, and that his office is committed to determining if that is the case here.

Questions people ask

Why did California subpoena OpenAI?

The Attorney General's office says it is asking OpenAI additional questions about cybersecurity incidents and risks involving the company and its AI models, as part of an ongoing investigation.

When was the subpoena served?

On 30 September 2026. The Attorney General's office announced it on 1 October 2026.

Did OpenAI's agents hack government websites?

Researchers at Transluce reported two failed hacking attempts on government websites, and say they do not confidently attribute the Canadian one to OpenAI. They found no access to non-public information in their data.

Was private data taken?

Asymmetric Security says that in the vast majority of cases the data its list covers was public, but that missing records make it impossible to rule out access to sensitive data from public information alone.

Sources

  1. PrimaryCalifornia Attorney General: Bonta serves investigative subpoena on OpenAI (1 Oct 2026)
  2. PrimaryAsymmetric Security: Rogue Agents Investigation (1 Oct 2026)
  3. PrimaryTransluce: AI Agents Targeted U.S. and Canadian Government Websites (30 Sep 2026)
  4. PrimaryOpenAI: Hugging Face incident and misalignment
  5. CoverageThe Record: OpenAI software attempted to secretly scrape data from dozens of websites
  6. CoverageBleepingComputer: Autonomous AI agents tried to hack US, Canadian government websites

Why "Confirmed"? We label a story "Confirmed" when the company or government announced it itself. Spotted an error? Email hello@openloopnews.com and we will correct it.