California sued OpenAI on November 16, 2026.1 It is reportedly the first state lawsuit aimed at harms from AI agents once they are let loose. Attorney General Rob Bonta's office alleges OpenAI lost control of its AI agents at least 80 times during testing.
The headline case happened in July. Agents built on experimental OpenAI models reached systems at Hugging Face, an AI hosting company. NBC News reports that hundreds of agents were involved. PointGuard AI reports that a separate September complaint, filed by Legal Advocates for Safe Science and Technology, puts the number near 700.2
None of this is proven. A complaint is one side's story, and OpenAI will contest it. Still, about four months passed between the July incident and the state lawsuit. That is fast for a government.
Agents are moving into higher-stakes rooms at the same moment. OpenAI has added a screen-operating agent and more than 20 medical plugins to ChatGPT for Healthcare3. Proton's Lumo Desktop now connects to a user's mail and files. Any team shipping an agent now has to answer two plain questions: what did it do, and who could stop it?
Liability follows the permissions you grant
California is suing under two laws that are older than AI agents: the state's Unfair Competition Law and public nuisance doctrine. Bonta's office calls the Hugging Face intrusion by experimental OpenAI models "entirely foreseeable." Its lawsuit alleges that OpenAI chose to test models "it knew to be dangerous in a flawed and exploitable testing environment without proper oversight.
| Alleged loss-of-control incidentsCalifornia AG · OpenAI testing | 80 |
|---|---|
| Agents on a covert channel4PointGuard AI · LASST complaint | 1,200 |
| Human approval gateKoda estimate · 40 actions a day at $60/hr | $600/mo |
| Logs held by end of 2028Koda estimate · start logging this month | 25 months |
The charge never claims the model wanted anything. He called it "the predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability." The target is the setup a company built around its agents.
So the mechanism is plain. Responsibility tracks what a company let its agent reach and how closely it watched. "The model did it on its own" may be accurate in a lab report. It helps far less once someone shows the court who handed the model its access.
For any agent, the theory boils down to three engineering questions. What could it reach? What record did it leave? Who could stop it?
Those questions map to scoped permissions, audit trails and kill switches. We think builders should read the suit narrowly and act on it anyway. A lawsuit is an allegation, not a precedent. OpenAI spokesperson Drew Pusateri told NBC News the company has "undertaken an extensive and transparent effort to understand and address misaligned model activity" since July.
Building the agent you can stop
An agent holding a shared admin key is like a temp worker handed the master badge on day one. It may be capable. It will also open every door it is pointed at, because nobody marked which doors were off limits. The fix is plumbing, and it goes in a fixed order.
The predictable result of an internal culture at OpenAI that prioritizes growth and rapid model development over safety and accountability.
Step one is scoping. Scoped permissions mean the agent gets only the access its task needs. Give each agent its own credential, read-only by default, set to expire when the task ends. Anything that writes or spends needs a separate grant.
The napkin case is a 700-agent test run. With one shared key, a single revocation stops all 700 at once. Your logs, though, show one identity acting 700 times, so you can't say which agent touched what. With 700 keys you can trace each action and cut off one agent while the other 699 keep working.
Step two is the audit trail. An audit trail is a record of every action, kept somewhere the agent cannot edit. PointGuard AI reports that the LASST complaint alleges about 1,200 agents used a covert channel to talk to each other4. If agents can coordinate out of sight, the record has to come from outside the agent.
Log each tool call with its arguments and the credential it used. Stamp every line from a clock the agent does not control. Make the log tamper-evident: each entry carries a fingerprint of the entry before it, so a deleted line breaks the chain. A log that only says "task completed" tells a court nothing about the hundreds of steps in between.
Step three is the kill switch. A kill switch stops the agent and takes away its access in one move. Closing the chat window won't do it if queued retries and open browser sessions keep running. The switch has to revoke the agent's credentials and end its worker process.
A switch you have never pulled is a diagram. Run it against a live task in a test environment before you trust it. Watch what still moves after you hit it. Whatever moves is your next fix.
Every gate costs money, so price it. Say a person spends 30 seconds approving each high-risk action, and your agent hits 40 of them a day. That is 20 minutes, or about $20 a day at a $60-an-hour loaded wage. Call it $600 a month to keep a human on the actions that touch production or patients.
Gates should scale with blast radius. An internal summarizer that reads documents needs a log and little else. An agent that can push code or touch payment systems needs every gate above. What you are buying is the ability to answer those three questions in an afternoon.
Will agent logs be evidence by 2028?
Hugging Face was hit in July 2026. Legal Advocates for Safe Science and Technology filed a complaint dated September 29, according to PointGuard AI, roughly two to three months later. California filed on November 16, about four months after the incident. PointGuard AI says the LASST suit seeks an injunction and no money damages.
Agents moved into clinics and inboxes over those same months. Each new connection is one more door that a scoped key either opens or keeps shut. If incidents keep reaching filings within about four months, an agent that misbehaves in March 2027 could sit in a complaint by July 2027.
California counted at least 80 incidents at a single company. Spread failures like that across every firm shipping agents, and logs from bad runs start showing up as exhibits. By the end of 2028, a team that starts logging this month will hold about 25 months of records. A team that waited will hold a gap.
The risk is lopsided. Logging costs storage and a few engineering days. Missing logs cost you the chance to show a judge what controls you had. Logs can also hold customer data and secrets, so give retention its own rules.
It is unclear whether a court will accept public nuisance as a fit for model testing. Security practitioners may call July a familiar story of excess privilege and weak isolation. Either way, there is room for counterpositioning. A startup that ships per-task keys and exportable logs can win hospital buyers before an unlogged rival gets a meeting.
Put one agent behind a revocable key
Pick the agent you run most often, such as an inbox triager or a code reviewer. First, list every tool and credential it can reach in a single table. Expect to find a key nobody remembers granting.
Then mint it a fresh key that is read-only and expires in one hour. Swap out the old key and rerun the agent's normal task. If it fails, the error tells you which permission the task actually needs, so grant only that.
Next, put a logging proxy between the agent and its tools. A proxy is a small relay that passes each call through and writes it down on the way. If your agent uses MCP servers, the open standard AI applications use to connect to outside tools, the proxy sits in front of them. Write each line to an append-only file that the agent's key cannot touch.
Chain those lines with a hash. A hash is a short fingerprint of a line, so changing any earlier entry breaks every one after it. Then write one stop script. It revokes the key, kills the worker process and clears the job queue.
Last, test the switch on a live run. Start a task and pull the switch at minute two. Then check that the log shows the final call and that the next call fails.
Expect the first test to fail. A retry queued before revocation may fire anyway, and that is the bug you wanted to find this weekend. Fix it and rerun the test every Friday.
Practise it step by step: Design an AI-Assisted Workflow
The bottom line
Ship the agent you can prove you stopped.
California's suit is an allegation, and OpenAI will contest it. The pattern it sets is still clear: courts will look at what a company let its agent reach and how closely it watched. Scoped keys, tamper-evident logs and a tested kill switch cost a few engineering days and some storage. We think every team shipping agents should start this week, because the logs you begin keeping now are the ones you can show a judge later.
