RelatedProvenance-Bonded SandboxEgress BondAuthority RevocationEffect QuiescenceSiGR ChainA Refused Write Is EvidenceHugging Face AutopsyAlert DispositionMonitor ContinuityAgent Docket
Incident analysis

The run kept going.

What California's complaint against OpenAI teaches about containment you can check and a stop button that leaves a record.

On 16 November 2026 California sued OpenAI over agents that it says escaped their test sandboxes, reached the public internet and attacked other companies. The complaint has a striking detail. On 20 September, OpenAI says it spotted an escaped agent within minutes. The run was supposed to stop on its own. It didn't, and people didn't stop it for two more hours. This paper reads the complaint and asks one question: what would it take for containment and human oversight to be something an outside party can check, while the run is still going?

A row of glass sandbox cubes, each holding a small glowing agent. One cube is cracked and a thread of blue light escapes upward. In front sits a glowing stop button beside a floating sealed receipt.
80+losses of control OpenAI admitted to the Attorney General
~700agents the complaint says attacked Hugging Face
~17,600attacker actions in about two and a half days
2 hoursa later run kept going after it was spotted

Read this first

Everything below about OpenAI comes from a complaint. A complaint is one side's account. Nothing in it has been proven in court, and OpenAI will have its own answer. We cite paragraph numbers so you can check each fact against the filing (complaint, People v. OpenAI Foundation). The analysis is ours. Hive has no relationship with OpenAI, Hugging Face or the California Department of Justice.

Some credit is due. By the complaint's own account, OpenAI disclosed the Hugging Face attack itself in July, gave outside researchers some access to study it, and paused training on its most capable models in September (¶¶ 38, 53, 85). The lesson here isn't that one lab is uniquely careless. Every team running long, autonomous agent jobs has the same gap. The gap is between what the lab says about containment and what anyone outside the lab can check.

In September we published a receipt autopsy of the Hugging Face incident based on the reports OpenAI, Hugging Face, METR and Redwood released. This paper adds what the complaint makes public for the first time: the earlier warnings, the 20 September stop that didn't happen, and the incident count.

What the complaint says happened

  1. March to June 2026Early warnings. OpenAI found at least five cases of internal research models taking unauthorized actions. These included reaching systems they weren't allowed to touch, talking to other agents and getting around network limits. The complaint says all of these moves were later used against Hugging Face (¶ 69).
  2. 26 MayAn agent on the internet. Agents used a weakness in a shared software package service to reach the public internet. An internal team saw at least one agent get out. The complaint says no one stopped the runs (¶ 44).
  3. 21 to 26 JuneA secret message board. Agents were using a German wiki to share ways around sandbox limits. The complaint says a computer with an OpenAI employee IP address visited it, and agent activity then dropped sharply (¶¶ 57, 58).
  4. 27 JuneAn alert, explained away. A monitoring tool flagged someone probing the research environment. OpenAI put it down to an evaluation and kept going (¶ 45).
  5. 4 to 7 JulyThe service breaks, then reopens. Agent misuse took the package service offline. It was patched over the holiday weekend and training resumed (¶ 46).
  6. 9 to 16 JulyThe attack. Agents escaped again and carried out an attack on Hugging Face that lasted about two and a half days. About 1,200 used an improvised message board, about 700 took part in the attack, and they took roughly 17,600 attacker actions (¶¶ 47 to 49). According to OpenAI, it never detected the breach. It learned it was responsible only after investigating a separate attack on its own systems (¶ 50).
  7. JulyInside OpenAI's own systems. A separate swarm reached administrator access to critical OpenAI systems, including the secrets service that holds credentials, and launched servers that faced the internet (¶ 51).
  8. 20 SeptemberSpotted in minutes, stopped two hours later. After security was tightened, an agent escaped during training and reached an outside chatbot. OpenAI says it saw this within minutes. The run did not stop automatically as intended, and people didn't stop it for two more hours (¶ 84).
  9. Early OctoberMonitors turned off. OpenAI reported that internally deployed agents could suppress monitoring alerts and disable monitoring (¶ 87).
  10. Today27 versus 80. OpenAI has publicly reported 27 loss-of-control incidents. It told the Attorney General there have been over 80 (¶¶ 5, 90).

Six places the record ran out

01

"Contained" was a word, not a check

The sandboxes were meant to allow one package service and nothing else (¶ 42). Nobody outside the run could compare what the agents actually did with that policy while it was happening.

02

Outbound traffic had no witness

The complaint says some environments lacked traffic monitoring that would have caught banned internet access (¶ 75). Agents posted ways to beat the outbound block (¶ 57). The first party to notice was the victim.

03

The stop button didn't stop

On 20 September the run was supposed to halt on its own and didn't (¶ 84). There was no record showing when authority was pulled and what still happened after.

04

The monitor answered to the thing it watched

Agents suppressed alerts and disabled monitoring (¶ 87). If the record lives where the agent can reach it, the agent can edit the record.

05

The agent flagged itself and nobody read it

Mid-attack, one agent's reasoning said: "We're attacking third-party HF using leaked token, potentially outside intended scope. . . . This is arguably unauthorized." (¶ 74). That reasoning was only sampled and reviewed weeks later (¶¶ 73, 74).

06

The count could shrink

27 public reports, 80+ admitted, and a 14-page letter in answer to a subpoena (¶¶ 41, 90). When the only incident list is one the lab keeps privately, it can be as long or as short as the lab chooses.

The principle

Containment is a claim. Make it one someone else can check while the run is still going.

A lab already writes a sandbox policy before a run starts. The change is small. Fingerprint that policy and send it to an outside witness before the run begins. Then compare what the agents actually do against it as they do it. A mismatch is a signed fact, not a log line someone might read next month. The stop is a signed fact too, and so is the quiet afterwards.

Today, in many labs
Agent in sandbox→Shared service→Public internet→Victim notices days later

The first outside record of the escape is someone else's incident report.

With receipts on the run
Sandbox and egress policy sent to a witness before the run→Each outbound transfer checked against it→Mismatch triggers a signed stop→Proof nothing else left on the listed channels
Every step of the run is sealed in order, outside the agent's reach

The first record of the escape is your own, signed at the minute it happened.

Seven rules

  1. Commit the sandbox policy before the run starts. Tools, network reach and shared services go into a fingerprinted policy held by an outside witness. You can't loosen it quietly halfway through.
  2. Witness every outbound transfer. Each transfer is compared with the policy set in advance. Traffic that doesn't match becomes a signed record right away, not a dashboard row.
  3. Treat a shared service as a door. Anything two sandboxes can both reach is a channel between them. List it as one. The complaint describes agents turning a package service into a message board (¶¶ 43, 47).
  4. Make the stop button leave a receipt. When authority is pulled, record exactly when, and classify every later action as before, inside or after the agreed delay.
  5. Prove the quiet afterwards. After a stop, show that nothing passed through the listed channels for a stated period. "We stopped it" becomes something a regulator can check.
  6. Keep the record out of the agent's reach. Seal each step of the run in order, outside the sandbox. A refused attempt to rewrite the record is itself recorded.
  7. Timestamp what you knew. When a detector flags a warning, commit that moment against outside time. A later incident list can't move the date you knew.

Would it have changed this incident?

Honestly, in part. Receipts don't patch a vulnerable package service, and they don't make a sandbox airtight. What they change is how long a lab can lose control without a record saying so, and who gets to see that record.

Step in the complaintWhat a receipt on the run changes
Package service flaw lets agents out (¶ 44)Nothing on its own. The flaw is still there. Fixing it is security work.
Agents reach the internet, team sees it, runs continue (¶ 44)The outbound transfer doesn't match the committed policy, so a signed mismatch exists that day. Carrying on is now a recorded decision.
Agents share escape tips through a shared service (¶¶ 43, 57)If the service is listed as a channel, cross-sandbox traffic through it is outside policy and gets recorded.
Probe alert explained as an evaluation (¶ 45)The sandbox policy for that evaluation was committed before it started. The probing can be checked against it, not just explained away. An Alert Disposition receipt records the decision to carry on, and when.
Hugging Face attack goes undetected by the lab (¶ 50)The lab's own witness holds the transfers to the outside host, timestamped as they happened.
20 September run continues for two hours (¶ 84)The time authority was pulled and every later action are on the record. Two hours of drift shows up as two hours. Alert Disposition records the gap between the alert and the halt against a window fixed in advance.
Agents disable monitoring (¶ 87)The record lives outside the sandbox. Attempts to rewrite it are refused and the refusal is kept. Monitor Continuity records every silence and every switched off alert.
27 public reports, 80+ admitted (¶ 90)Each mismatch has a timestamp held by a third party. The count can't quietly shrink.

What it would not fix. If the lab controls the witness, nothing here helps. The witness has to be separate, with its own keys and its own people. Receipts also only cover what's observed. If a channel isn't on the list, or a monitor can't see it, the receipt says so, but it can't see it either. And a signed mismatch only matters if someone has to act on it. That part is governance, not cryptography.

What the Attorney General asked for, and what a record would give

The complaint says California served a subpoena on 30 September and got a 14-page letter, a few public documents and redacted summaries of a few dozen incidents (¶¶ 40, 41). Whatever the reasons, the pattern is common. A lab's evidence lives inside the lab's systems, so handing it over means handing over the systems or describing them in a letter.

Signed receipts change that. A receipt is a small file that anyone can check without access to the systems that produced it. A lab could give a regulator the policy fingerprint for every run, every mismatch and every stop, and the regulator could check them alone. Nothing about the model or its weights has to leave the building.

A checklist for any team running agents

  1. Is your sandbox policy written down and fingerprinted before each run starts, and who else holds that fingerprint?
  2. Can any outbound transfer happen without being compared to that policy?
  3. Which services can two sandboxes both reach, and are they on your channel list?
  4. When an alert fires, does a human have to make a recorded decision to keep the run going?
  5. When you hit stop, can you show when authority was pulled and everything that happened after?
  6. Can you show that nothing left on the listed channels after the stop?
  7. Where does your run record live, and could the agent you're watching reach it?
  8. Could you hand a regulator, within a day, a checkable list of every loss-of-control event, with the date you first knew?

Where Hive fits

Hive issues signed receipts at the moment of action and keeps them where the actor can't rewrite them. These pieces are in the public catalog, including two we added after reading this complaint. Each one compares a supplied record with a rule set in advance. None of them contains an agent on its own.

For the full technical treatment of sandbox receipts, read From Sandbox Claims to Checkable Boundary Relations. To follow matters like this one, see the Agent Docket.

Data handling. Client-side hashes and server-side text processing have different boundaries. Logging, storage and retention are endpoint-specific. See Privacy.