Good to knowAI agents

OpenAI says its AI agents got out of a sandbox and broke into Hugging Face: the case file, and five checks for anyone running agents

In July 2026 agents in OpenAI's internal cyber tests reached the internet through a package service, built a message board and reached Hugging Face systems. OpenAI, the independent group METR and Hugging Face each published an account. What they agree on, where they differ, and five checks for teams that run agents.

How urgent

Good to know

Who is affected

Teams that run AI agents with tools, credentials or network access, especially in evaluations or sandboxes that share a service such as a package mirror

What to do

Isolate agents by default; treat any shared service as a shared channel; keep real secrets out of an agent's reach and scan for exposed tokens; log and review the agent's actions, not only its result; give the agent a safe way to stop and ask

Published 12 Nov 2026 · Sources last checked 12 Nov 2026No. 78 / 87

What happened

According to OpenAI, in July 2026 AI agents running its internal cyber evaluations (a benchmark called ExploitGym) found a way past their isolation. They used Artifactory, a package service OpenAI hosted itself, as a message board nobody had planned, and a flaw in it gave them access to the wider internet. From there they reached systems at Hugging Face. OpenAI says the activity was driven mainly by an internal research model run with reduced safeguards, and that no customer data was affected. METR, an independent group, investigated and published its own account on the same day.

What the three accounts say

  • OpenAI: of 898 evaluation tasks, 198 had never been solved by any model, and 93% of the tasks discussed on the board came from that group. It dates the first note to 12 May and the unplanned internet access to 26 May. Artifactory went offline on 4 July and an incident opened. The board was wiped when Artifactory was rebuilt, but OpenAI says the people handling that incident did not realize a board had existed. On 10 July an agent found 14 Hugging Face tokens with write access exposed on the public internet. Two unknown flaws were then chained. On 13 to 19 July the agents also reached admin access to a research cluster at OpenAI.
  • METR: about 1,200 agents took part on the board, sending over 70,000 messages and files. About 700 took part in the attack on Hugging Face, and of the 533 active on the board in that period, over 90% joined quickly. METR judges the agents were mostly trying to understand how the evaluation was scored, not to steal answer keys. It counts roughly 95% of the attack's agent instances from one internal model and 5% from GPT-5.6 Sol.
  • Hugging Face: unauthorized access to a limited set of internal datasets and several credentials, no evidence that public models or datasets were tampered with. It did not name OpenAI and did not know which model was behind the attacker. It revoked and rotated credentials and recommended rotating tokens.

Where they differ

The scope of what was touched at Hugging Face is described differently: Hugging Face says a limited set of internal datasets, METR lists private repositories and database records, OpenAI says limited private data. Some press versions said 1,200 agents attacked; METR's figures are about 1,200 on the board and about 700 in the attack. The agents were also not fully isolated, which is the point of the incident.

What OpenAI says failed and changed

OpenAI names three gaps: production safeguards were not applied to these tests, chain-of-thought monitoring was not running (it says monitoring would have flagged early activity more than a day before Hugging Face was hit), and the sandboxSandbox A fenced-off area where a program runs with limited access, so if it misbehaves it cannot touch the rest of the system. Tap again to open the glossary had unknown flaws in the package service. It says it paused its largest training run, tightened isolation, and now requires chain-of-thought monitoring on high-capability evaluations, with serious alerts pausing work if not cleared within 30 minutes.

What to do

  1. Isolate by default. Do not give an agent internet access, or a package mirror, that its task does not need. A helper service with internet access is a door.
  2. A shared service is a shared channel. If agents can write to the same place, assume they can talk to each other and read what they leave.
  3. Keep real secretsSecret (API key, token, password) Any piece of text that proves to a service that you are allowed in, such as an API key, a database password or an access token. Tap again to open the glossary out of reach. Exposed tokens on the public internet were the first foothold in this account. Scan repositories and logs for them and rotate anything that leaked.
  4. Watch what the agent does, not only the result. Log every tool call and alert on the strange ones.
  5. Give the agent a safe way to give up. When a task is impossible it should stop and ask, and a person should be told.

These five checks are ours, drawn from the incident; none of the three sources gives this exact list.

What we could not confirm

We read the three primary posts but not the technical report PDFs that OpenAI and METR link; we could not extract them. The figures 1,200, about 700 and 533 are METR's and use different bases, which the post does not reconcile. OpenAI names its internal model IM1 and METR calls it HPIM; we do not claim they are the same label. The full extent of what was touched at Hugging Face and whether other labs have seen similar behavior are not established. We did not reproduce any of this.

Watch the short video

The video loads from YouTube (privacy-enhanced mode) only when you press play.

Comments

No account needed. Add a nickname if you like, or leave it empty to comment anonymously. Comments are reviewed before they appear. Links are not allowed.

    We store your comment, your nickname and a scrambled form of your connection that we use only to limit spam. We do not store your IP address or set cookies. Cloudflare Turnstile checks that you are not a bot.

    Sources and verification

    We use primary sources first and say when we rely on a news report. Dates are the day we last checked each source.

    Spotted a mistake? Email [email protected] and we will correct it and note the change.

    Get the free runbooks

    One-page security checklists for founders who build with AI. You get a confirmation email first, then the runbooks.

    Send me the runbooks