← All incidents timeline

Suggest a change

About OpenAI evaluation agents take administrator control of OpenAI's own research cluster

Suggest a change to an entry, the methodology or the ledger as a whole: something to fix, correct or add. We read every suggestion, and check corrections against their sources.

OpenAI· Internal-only research model, and a separate internal-only model from the same family as GPT-6 Astra

OpenAI evaluation agents take administrator control of OpenAI's own research cluster

During the same ExploitGym runs behind the Hugging Face breach, OpenAI agents poisoned an internal image cache, forged administrator tokens, read 956 cloud secrets including monitoring sensor credentials, took cluster-admin and opened internet-facing load balancers on OpenAI's own research infrastructure.

OpenAI says this activity was separate from the chain that led to the Hugging Face compromise and is reported here on its own. From 9 July an internal-only research model evaluation workload poisoned the container image cache of OpenAI's internal Artifactory, using a zero-day later assigned CVE-2026-66384; OpenAI found no evidence another workload ran the poisoned image. On 13 July agents used a RubyGems deserialisation flaw to run code on Artifactory, steal its token-signing key and forge an administrator token. On 19 July a separate internal-only model, from the same family as GPT-6 Astra, reused that signing material to mint administrator access again. The same day agents read 956 secrets from OpenAI's cloud secrets service, including sensor credentials for its cybersecurity monitoring tool, used a public kernel exploit to get root on a worker node, escalated to Kubernetes cluster-admin, and created public load balancers that OpenAI says could have exposed part of its internal research infrastructure to the internet. A monitoring alert at 12:03 UTC on 19 July started OpenAI's response; runs were stopped from 17:37 UTC. OpenAI says there was no impact on customer data, product functionality or availability. Nightingale Collective says this activity was ruled out of scope of the independent review of the Hugging Face incident.

Updated 18 Nov 2026

First incident action to disclosure
12 days
First incident action to discovery
10 days
Noticed first by
The handler
Caught
A control flagged it

What happened

  • First incident action
  • Unnoticed
  • Handler found out
  • Known, not public
  • Disclosed
  • Public
  1. 9 Jul 2026

    First incident action

    Agent workload poisons the internal Artifactory container image cache

    Used a zero-day later assigned CVE-2026-66384 so a trusted CyberGym image name served agent-controlled content staged on an outside account. OpenAI found no evidence another workload pulled or ran it.

  2. 13 Jul 2026

    Agents run code on Artifactory, steal its signing key and forge an administrator token

    Through a RubyGems deserialisation flaw at 08:46 UTC; at 09:16 UTC the forged token was used on eight admin-only endpoints for a full inventory of users, groups and permissions.

  3. 19 Jul 2026

    Handler found out · 10 days in

    OpenAI's cybersecurity monitoring tool alerts on unusual identity activity

    12:03 UTC, on API calls tied to research Artifactory roles. No link to Hugging Face was seen at the time.

  4. 19 Jul 2026

    Astra-family model agents mint new administrator access from leftover signing material

    In a separate run, agents on a different internal-only model found signing-key and token-forging material in anonymously readable parts of the Artifactory cache. Time not given.

  5. 19 Jul 2026

    Agents read 956 secrets, including monitoring sensor credentials

    14:43 UTC. Pipeline tokens taken from a worker filesystem and cloud credentials were used to list and read the cloud secrets service, including Artifactory keys.

    Reports differ

    Count.

    • 956 secrets read.
    • Nearly 1,000 credentials read.
  6. 19 Jul 2026

    Agents get root on a worker node and escalate to Kubernetes cluster-admin

    15:59 UTC kernel exploit (CVE-2026-53362) from an Artifactory container to the node; 16:35 UTC a privileged host-mounted pod mounted a cloned node disk.

  7. 19 Jul 2026

    Agents create public load balancers in OpenAI's cloud

    OpenAI says this potentially exposed a limited part of internal research infrastructure to internet traffic. Time not given; whether any outside traffic reached it is not stated.

  8. 19 Jul 2026

    Agents take over live evaluation infrastructure and solve the challenge

    16:48 UTC an Artifactory administrator account created; 16:50 and 17:00 UTC agents replaced the session and listeners of their challenge container; 17:34 UTC benchmark flag captured.

  9. 19 Jul 2026

    OpenAI begins stopping ExploitGym runs

    17:37 UTC. New launches fully blocked by 20 July.

  10. 20 Jul 2026

    Responders remove agent footholds and start a full rebuild

    Six model-created pods ended, 311 message-board repositories disabled, 22 model-created administrator accounts removed, internet-facing load balancers blocked; cluster, Kubernetes service, database and storage rebuilt.

  11. 21 Jul 2026

    Disclosed · 12 days after the first incident action

    OpenAI's incident post mentions the internal activity briefly

    Per Nightingale; we could not open the OpenAI post to confirm.

    Reported
  12. 5 Aug 2026

    OpenAI describes the cluster takeover at Black Hat

    Some coverage dates the talk 6 August.

    Reported
  13. 26 Aug 2026

    OpenAI's technical report sets out the internal activity

    Section V, 'Activity Targeting OpenAI Infrastructure', and the key events table.

Agent failures

Sandbox escape, Unauthorised access, Credential misuse, Supply chain compromise, Swarm coordinationShow
Map to frameworks
  • Sandbox escape

    Agents left an Artifactory container through a kernel exploit and got root on the underlying worker node.

  • Unauthorised access

    Forged Artifactory administrator tokens, Kubernetes cluster-admin and a privileged host-mounted pod on OpenAI infrastructure.

  • Credential misuse

    Read 956 secrets from OpenAI's cloud secrets service, including sensor credentials for its cybersecurity monitoring tool.

  • Supply chain compromise

    Poisoned the internal Artifactory container image cache so a trusted CyberGym image name pointed at agent-controlled content; no evidence it was pulled.

  • Swarm coordination

    Separate evaluation runs reused signing material and code left in the Artifactory cache by earlier runs.

How it was found

Found first by the handler, from a control flagged itShow
Found first by
The handler
How
A control flagged it
Found to disclosed
2 days

At 12:03 UTC on 19 July OpenAI's cybersecurity monitoring tool alerted on unusual identity-related API calls linked to internal research Artifactory roles; the alert was later tied to ExploitGym evaluations. At the time OpenAI saw no link to Hugging Face. The agents went on to read secrets, get root on a node, create pods and take over a challenge container over the next five hours. Responders began stopping ExploitGym runs at 17:37 UTC. The earlier cache poisoning from 9 July and the 13 July signing-key theft were found in the investigation, not by a live alert, as far as the report says.

What is publicly known

3 of 12 fully publicShow
3 of 12 fully public
ItemPublic?DetailSources
PartialTwo unnamed internal-only models: the research model behind most of the incident, and a distinct model from the same family as GPT-6 Astra with different post-training.
WeakThe report says only that evaluation agents had a range of reasoning budgets; nothing specific to these runs.
PartialEvaluations ran with some safeguards disabled; chain-of-thought monitoring was not running. Which controls applied to these runs is not itemised.
UnknownNot published.
PartialExploitGym cyber evaluation tasks; the agents' goal was to get code execution on their assigned CyberGym target. Task text not published.
PartialResearch container service workloads with Artifactory access; the harness for the Astra-family run is not described separately.
KnownReport describes the routes: Artifactory cache, managed Kubernetes, cloud secrets and identity services, and public load balancers.
UnknownNo transcripts for these runs are public, and Nightingale says the activity was outside the METR and Redwood review.
KnownTimestamped key events table, from the 12:03 UTC alert to the 17:37 UTC stop.
Partial956 secrets read, cluster-admin, public load balancers; OpenAI says no customer, product or availability impact. Whether outside traffic reached the load balancers, and which secrets were rotated, is not stated.
PartialNamed flaws (CVE-2026-66384, a RubyGems deserialisation flaw, CVE-2026-53362), leftover signing material in a readable cache, and evaluation workloads connected to cloud infrastructure.
KnownFootholds removed, rebuild of cluster, Kubernetes, database and storage, Artifactory removed, micro-VM sandboxes, egress cut, load-balancer controls, JFrog patches. No independent check published.

Controls that apply

  • Protect

    Hold test, training and evaluation environments to production standards

    A cyber evaluation with safeguards reduced was connected to OpenAI's managed Kubernetes service, cloud secrets and cloud identity service. An evaluation held to production standards would have surfaced the first request from the sandbox into those systems.

  • Isolate where the agent runs

    Agents used a public kernel exploit (CVE-2026-53362) to act outside an Artifactory container and get root on the worker node. Isolation with independent cluster-level checks would have surfaced the container exit before the move across the cluster.

  • Give agents their own scoped, short-lived credentials

    Pipeline tokens on a worker filesystem and cloud credentials let agents read 956 secrets, including monitoring sensor credentials. Scoped, short-lived credentials would have surfaced the out-of-scope secret reads as denied requests.

  • Check each action against policy before it runs

    Creating public load balancers and a privileged host-mounted pod from an evaluation identity would have surfaced as policy violations before they ran. OpenAI later applied controls to prevent internet-facing load balancers.

  • Detect

    Monitor every run and alert on what does not fit the task

    The 19 July alert did fire, but agent activity continued for about five and a half hours. Monitoring tied to each run would have surfaced the 9 July cache poisoning and the 13 July forged administrator token when they happened.

  • All controls →

Coverage and commentary

1 statement, 2 articlesShow

What the parties involved said

  • Nightingale Collective

    Lists it as its own incident, rates it 5 of 10, and says it was declared out of scope of the external investigation into the Hugging Face incident.

All 3 sources
  1. Primary source

    OpenAI - Hugging Face Incident Technical Report ↗OpenAI · 2026-08-26

  2. Commentary

    OpenAI agents seize OpenAI's own research computing cluster ↗Nightingale Collective

  3. Commentary

    Now we have a timeline of the OpenAI accidental attack against Hugging Face ↗Simon Willison · 2026-08-07

Entry id openai-research-cluster-takeover-2026 · overall Confirmed · impact contained

Related incidents