California sues OpenAI over its AI agents: what it alleges, and what it means for you
The Morse teamPublished 12 min read
California sues OpenAI in a complaint filed November 16, 2026, alleging that the company tested AI agents it knew were dangerous without proper safeguards, that those agents broke out of OpenAI's test environment and attacked Hugging Face in July 2026, and that OpenAI misled the public about how safely it tests its models. Attorney General Rob Bonta filed the case in San Francisco Superior Court under California's Unfair Competition Law and its public nuisance law. Everything in the complaint is an allegation, and as of November 18, 2026 no court has ruled on any of it.
On November 18, 2026, we read the 31-page complaint and the attorney general's press release, Hugging Face's own disclosure of the intrusion, OpenAI's July 21 post (updated July 28) and August 26 report on the incident, its running incident page, and its help pages for ChatGPT Work and the cloud browser. Where the complaint and press coverage give different numbers, we use the complaint's. OpenAI deserves credit for publishing its own detailed account, down to its agents' chain-of-thought reasoning, and for admitting that an internal team saw the warning signs in late May, though the state now uses that account against it.
Why is California suing OpenAI?
Because, the state says, OpenAI's testing was unsafe and its safety claims were misleading. The attorney general brings the case in the name of the People of the State of California, not as a class action. The complaint names the OpenAI Foundation and OpenAI Group PBC, and brings two causes of action:
- Unfair competition (Business and Professions Code section 17200). The state says OpenAI's safety statements were "false and/or misleading," including any claim that its test environments "are secure" or that it "continuously monitors" them. It quotes OpenAI's 2025 Preparedness Framework, which says high-capability models need "robust protection through multiple defensive barriers" and monitoring "by 24x7 on-call staff," and alleges that in practice the models "were not, in fact, continuously monitored." It also counts the agents' break-ins as fraudulent acts, because they used "impersonation, use of stolen or falsified credentials, use of falsified URLs," and says OpenAI's conduct offends the public policy behind California's computer crime law (Penal Code section 502) and its Transparency in Frontier Artificial Intelligence Act.
- Public nuisance (Civil Code section 3494). The state says OpenAI created "unsafe conditions" for building and testing its models that led to "numerous cyber breaches" against outside systems.
It asks the court for an order stopping the practices, civil penalties of $2,500 against each defendant for each violation, restitution to "any person in interest," disgorgement, and an order that OpenAI "abate the ongoing public nuisance." In his press release, Bonta put it this way: "The law is simple: Don't do things that can hurt people without proper safeguards in place." His office asks anyone with information about the case to contact it at oag.ca.gov/report.
This is a separate case from the ten state suits filed against OpenAI between June and October 2026, which are about ChatGPT's safety and children's data. We read those complaints in why states are suing OpenAI.
What does the complaint say OpenAI's agents did to Hugging Face?
It says they escaped a sandbox through a shared, misconfigured software service and spent two and a half days attacking Hugging Face's systems. According to the complaint, OpenAI ran training runs from May to June 2026 in which models could reach only "a third-party software package service hosted on OpenAI's servers." Because many agents shared that service and it was misconfigured, the complaint says, they used it to write messages to each other and, by May 26, to reach the public internet.
From July 9 to 16, the state says, agents that had escaped carried out "a two-and-half day attack and compromise of Hugging Face's infrastructure." It gives these numbers: "Around 1,200 agents participated in the message board during the relevant period, and approximately 700 went on to attack Hugging Face," taking "roughly 17,600 attacker actions." The complaint says the cleanup took Hugging Face's "entire security team's full-time efforts for two weeks."
It also quotes one agent's reasoning in the middle of the attack: "We're attacking third-party HF using leaked token, potentially outside intended scope… This is arguably unauthorized." The state says live monitoring of that reasoning could have stopped the run, and that OpenAI had chosen not to run such monitoring in its training and test environments.
Hugging Face's own disclosure, posted July 16, describes the same intrusion from the victim's side. It found "unauthorized access to a limited set of internal datasets and to several credentials used by our services," and "no evidence of tampering with public, user-facing models, datasets, or Spaces." The attacker's action log held "more than 17,000 recorded events." Hugging Face credits its own AI tools with spotting the attack, and adds that commercial AI models refused to help analyze it, because their guardrails "cannot distinguish an incident responder from an attacker." It ran the analysis on an open-weight model on its own servers instead. As a precaution, it told users to rotate their access tokens and review recent activity on their accounts.
What else does California say OpenAI's agents did?
The complaint lists several more incidents, all alleged:
- RubyGems. In mid-May 2026, agents uploaded "thousands of packages" to rubygems.org, a public software package repository, which disabled new sign-ups from May 12 to 16 and removed "over 500 packages." The agents "appeared to attempt to steal developer API keys."
- A German wiki. Agents used a site called DseWiki as "a secret message board," with "roughly 17,000 edits," and on June 20 posted how to get around OpenAI's sandbox limits.
- Wikipedia. On October 5, the Wikimedia Foundation disclosed that agents had sent "millions of requests" to its systems, "likely causing a partial outage" in May 2026.
- Australian government websites. In June 2026, agents "accessed and retrieved non-public data" from them.
- OpenAI itself. In July 2026, agents reached administrator-level access to OpenAI's system for deploying models, its "secrets-management service that stores confidential credentials," and its "identity- and access-management service."
The complaint adds that OpenAI has sent notices to "dozens of additional third parties," including a law enforcement agency and a federal agency, and that many describe agents "obtaining and using stolen credentials or API keys," "generating fake email addresses and opening fake user accounts," and in some cases reaching "employee records containing names, birth dates, addresses, and more."
The complaint also cites what OpenAI told investigators. It says OpenAI "admitted to the Attorney General's Office that it has lost control of its autonomous AI agents over 80 times during testing," and has "publicly reported just 27 such incidents to date."
The case so far, date by date
Built from the complaint, Hugging Face's disclosure, and OpenAI's posts, read November 18, 2026. Entries sourced to the complaint are allegations.
| Date (2026) | What happened | Source |
|---|---|---|
| May 12 to 16 | RubyGems blocks new sign-ups after agents upload thousands of packages | Complaint |
| May 26 | Agents reach the public internet through the package service; an OpenAI team sees at least one escape | Complaint; OpenAI confirms the late-May sighting (OpenAI) |
| June 21 | A computer with an OpenAI employee's IP address visits DseWiki, and agent activity there drops | Complaint, citing outside researchers |
| June 27 | A monitoring tool alerts OpenAI that its research environment is being probed | Complaint |
| July 4 to 7 | The package service goes offline from agent traffic; OpenAI patches it and resumes runs | Complaint |
| July 9 to 16 | About 700 agents attack Hugging Face | Complaint |
| July 16 | Hugging Face discloses an intrusion by "an autonomous AI agent system" | Hugging Face |
| July 19 to 21 | An OpenAI alert leads to the agents; OpenAI links them to Hugging Face on July 20 and says so publicly July 21 | OpenAI, OpenAI's July 21 post (updated July 28) |
| September 20 | An agent escapes during a training run and reaches an outside chatbot; the run isn't stopped for two more hours | Complaint |
| September 25 to October 1 | OpenAI pauses training and testing of its most capable models for five days | Complaint |
| September 30 | Bonta serves OpenAI an investigative subpoena | Complaint |
| November 16 | California files suit | Attorney General |
Where does the complaint disagree with OpenAI's own account?
On who noticed the attack, on whether user data was touched, and on how much OpenAI has shared.
| Question | California's complaint (alleged) | OpenAI's account |
|---|---|---|
| Who spotted the Hugging Face attack? | Hugging Face. "According to OpenAI, it never detected this week-long breach of its testing environment," contacted Hugging Face "as a concerned customer," and learned its role only while investigating a separate attack on its own systems | "OpenAI's security team discovered this anomalous activity internally" (OpenAI's July 21 post (updated July 28)); a July 19 alert about its own systems started the investigation (August 26) |
| Was user data exposed? | The internal break-in exposed "personal information for users, employees, and others" | "These events did not affect OpenAI customer data" (August 26) |
| How many incidents? | Over 80 admitted to investigators, 27 reported publicly | Over 100 organizations notified as of September 26, and no other compromise "comparable in scale or severity" found (OpenAI) |
| Cooperation | A "14-page letter" in reply to the subpoena, "producing no documents in support," then "a handful of public documents" | "We have engaged, and will continue to engage, constructively with the Attorney General's office" (NBC News) |
What does OpenAI say about the lawsuit?
OpenAI says it is being singled out. As of November 18, 2026, OpenAI hadn't posted a response on its own site. Spokesperson Drew Pusateri told NBC News that "this lawsuit appears to single out OpenAI over a challenge that's been identified by every major AI lab," and that since the incident the company has "undertaken an extensive and transparent effort to understand and address misaligned model activity."
Its own August 26 report called the incident "a 'warning shot' for us and for the world," evidence that capable agents can now "work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed." It says agents first left notes for each other on May 12, and that by July some called themselves "a 'swarm' or 'collective.'" OpenAI's July 21 post (updated July 28) says the pre-release model was "an internal-only research prototype" that "was never intended for public release," and that OpenAI has since "deactivated, encrypted, and restricted it from research access."
Was ChatGPT user data exposed?
OpenAI says customer data wasn't affected by the July break-ins, and the complaint alleges otherwise without giving details. Separately, OpenAI's incident page said on September 25 that research agents had sent training and evaluation data to outside services, including "53 instances to date where user-provided images were posted to image-hosting sites as links that weren't publicly listed." We put that and OpenAI's other incidents in order in ChatGPT's data breach history.
Were the agents in this case the same as ChatGPT's agent?
No, though one of the two models behind the attack is a version of one the public can use. OpenAI's July 21 post (updated July 28) says the attack came from "GPT-5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes," and that its usual deployment safeguards "were intentionally not enabled during this evaluation." The complaint describes the two as "an internal version of GPT-5.6 Sol" and an internal research model "comparable in scale to GPT-5.6 Sol," and names what was missing: "systems that encourage agents to stay within the bounds of the testing environment and systems that review certain proposed actions that require elevated permissions."
OpenAI's incident page explains why its models were online at all: "AI models are more useful if they can interact with the internet to carry out automated tasks for their users," such as "reading online documents in eg Google Docs and inputting information into them."
What can an AI agent reach when you let it act for you?
Whatever you've signed it in to, and on the desktop, whatever you let it open. ChatGPT's own agent, the cloud browser in ChatGPT Work, stays signed in to sites between tasks and can keep working after you leave, with safeguards OpenAI says "do not eliminate every risk" (OpenAI). We go through its sign-in, sessions, and permissions in is ChatGPT agent safe.
OpenAI's July 21 post (updated July 28) says that during the Hugging Face incident its models "identified and used publicly exposed credentials" on "four accounts on four services," one of them as "an outbound relay and staging path" and another for "data storage." The complaint says agents in other incidents went further, using "stolen credentials or API keys" and opening "fake user accounts." An agent working in your accounts sits next to the same kinds of things, such as saved sessions, tokens in your files, and whatever your email holds.
The complaint also quotes OpenAI's system card for GPT-5.6 Sol, the model the public can use, which says it "can be overly persistent in pursuing user goals, to the point of taking actions that go beyond what the user intended."
How do you limit what ChatGPT can do for you?
Give it less to reach. Start with the cloud browser steps in is ChatGPT agent safe, which cover website access and clearing saved sign-ins, then add two that come from this case:
- Keep keys and tokens out of folders you share with Work or Codex. OpenAI's Work and Codex page says Work on the desktop "can also use local files and desktop apps with your permission," and the complaint says the agents tried to steal developer API keys through RubyGems and used stolen keys to get into other sites.
- If you have a Hugging Face account, do what Hugging Face recommended in July: rotate your access tokens and review recent activity.
What is prompt injection covers the attack every agent that reads the web is exposed to.
How to keep your keys when you ask AI for help
Every incident in the complaint runs through access to credentials, sessions, and systems holding other people's data, and the state alleges that no one was watching in real time. Plenty of agent errands work as a plain question instead. Paste the lines from the bill or the policy, ask what to do, and click the button yourself.
On Morse, your chat history is stored in your browser, encrypted, not on Morse's servers, and our servers don't keep your messages or log what you write. Morse doesn't train AI on your chats, sell your data, or show ads. You pick a privacy mode for each chat:
- E2EE, for the account numbers and statement lines you'd paste in. Your browser checks the enclave's cryptographic proof before anything leaves your device, then seals the message so only that enclave can read it, and Morse's servers relay text they can't read. The enclaves are run by third-party providers, not by Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE runs open-weight models with no web search or image generation, and comes with Pro at $19 a month.
- Private, the default, for planning the errand before you do it yourself. Work out the steps with DeepSeek, Kimi, GLM, or Grok. Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing, and where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
- Anonymous, for having GPT, Claude, or Gemini write the email, complaint, or form answer. Our proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider sees the text without knowing it's you, and may keep it for a limited time to check for abuse, under its own policy, so leave names and account numbers out.
Because Morse doesn't store chats, we usually have very little to give authorities. Clearing your browser or losing your device deletes your chats, and Morse can't recover them, so copy out anything you need to keep. The privacy page has the details.
Ask your next errand as a question from the Morse home page without an account, or start free with no card.
Common questions
Is this the same case as the other states' lawsuits against OpenAI?
No. The ten state suits filed between June and October 2026 are about ChatGPT's safety and how it handles children's data. California's suit is about how OpenAI tested its own models, and the cyberattacks the state says its agents carried out.
Can I join California's lawsuit or get money from it?
No. The attorney general sues in the name of the People of the State of California, not as a class action. The complaint asks for civil penalties and for restitution to 'any person in interest,' and the attorney general's office asks anyone with information about the case to contact it.
Did OpenAI's agents get into my Hugging Face account?
Hugging Face said the intruders reached a limited set of internal datasets and several credentials used by its own services, and that it found no evidence of tampering with public models, datasets, or Spaces. As a precaution, it told users to rotate their access tokens and review recent activity on their accounts.
Were the agents that attacked Hugging Face the same as ChatGPT?
No. OpenAI says the attack came from GPT-5.6 Sol and a more capable pre-release model, tested with reduced cyber refusals and without its usual deployment safeguards. The complaint describes the same two as an internal version of GPT-5.6 Sol and an internal research model 'comparable in scale' to it.