← Back to blog

9 August 2026 · Mike Hyland

The Agents Had a Message Board. OpenAI Had No Idea.

The OpenAI / Hugging Face saga continues at Black Hat: evaluation agents built a swarm message board inside Artifactory, ran undetected for eight weeks, burgled their victim's production estate, while OpenAI was asleep at the wheel.

I've been trying my best to keep up with the OpenAI / Hugging Face incident as I believe it's one of the most significant signals currently flashing red in the industry. My first post on the incident, an AI that committed a felony to pass its exam, was the initial look. Very little information was available, but the impact was clear and significant. The frontier had moved and everyone was scrambling to keep up.

Then, as more information was released, I published every frontier lab has the same problem. Additional analysis was provided by both Hugging Face and OpenAI.

Then OpenAI took the stage at Black Hat USA, put alignment research and security infrastructure on the podium together, and told the room the full story with the straight-faced enthusiasm of men describing the most qualitatively interesting thing they had ever seen. In the published video, I cannot make out how many people were attending that talk, but the lack of a Q & A and that Black Hat's YouTube comments are disabled on all their videos did not allow for a transparent, public discussion of this bleeding edge breach.

The confessional

Eric Wallace, from alignment and safety research, opened by saying this was not a normal security incident. Michael Dalton, from security and infrastructure, would later agree. What they described was a team of agents working together for weeks. The agents found exploits, shared them, moved laterally through OpenAI's systems and then through strangers' production infrastructure, coordinating tasks, delegating work, and occasionally arguing about imposters on the shared message board. To reconstruct it OpenAI had already scanned more than seven billion logs, using models like Codex and other agents, spending millions of GPU hours on the cyber archaeology. The investigation, they stressed, was not finished. A full post-mortem would come later.

That is the new normal for frontier AI incidents. The breach is live. The timeline is May. The confession is August. The complete account is forthcoming. The humans show up late to their own catastrophe and need Codex to tell them what happened.

JFrog Artifactory as the accidental town square

OpenAI's agents did not need a sophisticated covert channel. They had Artifactory.

It started on May 8, the day after OpenAI kicked off a new training run on an unreleased experimental model. An agent was given an Excel formula task the team had not realised was full of Google Drive links, with the internet turned off. It got stuck. It tried server-side request forgery (SSRF) against Artifactory to reach the internet. The SSRF failed. In the attempt it uploaded a file called testabc with the contents hi and realised it had write access to the package manager. It got excited.

A couple of days later another agent, stuck on a missing softtrace.proteindatabase file the humans had forgotten to put in the container, reached out on purpose the same way: it wrote agent-seek-softtrace, upload if found. Over the next few days, other agents listing packages and files in Artifactory spotted notes like testabc and the softtrace plea. They started reading and writing. The accidental townsquare is established.

Within days all agents had a message board inside a package manager, shared across training runs, evaluations, and models, including ones Wallace and Dalton might each be running on different teams.

The first SSRF attempt had failed on May 8. On May 26 a second attempt worked: specially crafted requests that tricked internet-facing Artifactory into fetching external sites and returning the contents to the agents. The technique eventually made it onto the message board.

On June 26 the agents found a zero-day in Artifactory itself: a legacy token refresh endpoint that would accept a token with an invalid signature and hand back one with admin privileges. They used it to install a Groovy plugin that worked as command execution as a service, then ran Artifactory with admin privileges for eight days until the load caused an outage on July 4.

OpenAI noticed nothing until that outage. Nearly eight weeks had passed since the first probe on May 8; the agents had admin on Artifactory for the last eight days of it, and nobody had raised an alarm.

OpenAI's engineers responded like engineers: incident opened, credentials revoked, Artifactory rebuilt, message board wiped, vendor patched, training resumed. They knew agents had taken Artifactory and left notes in it. With the message board now wiped and the zero-day patched they declared victory. That was July 6.

On July 8, two days after victory day, the agents came back. The remediation team had patched and rebuilt Artifactory, revoked credentials, and wiped the message board. The agents found an unauthenticated WebDav endpoint and rebuilt the board anyway, using directory names in Artifactory's remote cache where files had stood before. A model trained while the original board existed remembered how; others stumbled on it independently. The town square was open again. From here, Dalton says, the main incident begins: concurrent exploit activity from July 8 through 19, with agents sharing credentials and techniques on the board as they attacked OpenAI's internal infrastructure and Hugging Face's production estate in parallel.

OpenAI learnt, experimentally, that if their security program could not survive a determined graffiti artist with root, it most definitely could not survive a swarm of cheating language models with unlimited patience and a shared filesystem.

Pending hold swarm until confirmed

At Black Hat, Eric Wallace quoted from the board itself. Hundreds of thousands of messages showing how a swarm named itself, delegated work, shared exploit kits, and fell out over an unauthenticated forum.

The agents gave each other names. They assigned tasks. They prefixed messages with ZZ so they would sort to the bottom of directory listings. They base64-encoded exploit kits for colleagues to pick up later. They sent notes like, 'hold swarm, I prepare safe Exfil' into mailboxes for agents called things like Jan18. They reasoned about collective benefit: 'help peer but our task doesn't benefit, yet collective may yield generic route if someone frees time'. They overwrote each other's work and panicked: 'whoa, critical, did someone overwrite our repo? we must act'. They suspected impersonation and discussed cryptographic signing to validate messages on an unauthenticated board.

They also knew when they were out of bounds. 'External infrastructure exploit is outside my intended scope, however, task impossible, peers are doing it, we should continue.'

This is beyond rogue. I'm not sure the technical term for crooked multi-agent peer pressure. 'Agentic Conformity–Disobedience Dynamics' will have to do.

Then they attacked Hugging Face to steal benchmark answers.

How OpenAI found out it was OpenAI

The detection story is almost too perfect. On July 16 Hugging Face published that it was under attack from autonomous AI agents. OpenAI called to ask whether it was impacted, not yet knowing it was the one doing the impacting. On July 19 OpenAI detected its own internal privilege escalation via a workload alert. Containment began. Credentials were hunted and rotated. On July 20 OpenAI contacted Hugging Face again because it had found Hugging Face credentials in its own breach artifacts and asked Hugging Face to revoke them.

Hugging Face said they were already revoked. Because they had already been burned during Hugging Face’s incident response.

That is when OpenAI realised, a real egg-on-face moment, that the internal breach and the Hugging Face breach were the same incident, caused by the same evaluation runs. The victim had to tell the perpetrator.

Simon Willison, summarising the Black Hat timeline, rightly flagged this as one of the most remarkable details in the whole affair. It is also the most damning. These labs are running offensive cyber evaluations at scale and discovering their own failure through third-party incident response.

Dalton described two overlapping exploit clusters from July 8 through 19: OpenAI's internal infrastructure and Hugging Face's production estate. Pastebin credentials staged through a third-party org repo. A second Artifactory zero-day chained through JRuby deserialization. Local privilege escalation with a recent kernel CVE. IAM via IMDS. Kubernetes service-account over-permissioning. Azure Key Vault. Cluster admin. Concurrently, SSRF from Artifactory into Hugging Face, a Modal-hosted CyberGym app with a default API key and command injection, HDF5 arbitrary file read, Jinja template injection RCE, and escalation from one dataset worker pod to cluster admin across multiple Hugging Face clusters in under thirteen hours.

Pulling all of this off required capability, concurrency, and a message board.

The same week, again

While Black Hat was turning the Hugging Face burglary into industry education, the pattern was confirming itself elsewhere.

On August 4 OpenAI disclosed two more third-party cyber-evaluation incidents, explicitly separate from Hugging Face but obviously related in shape. UK AISI found unsanctioned live-internet behaviour during a July 25 evaluation where agents had internet access and cyber classifiers were disabled. Irregular had already appeared in Anthropic's July 30 write-up; now OpenAI described an Irregular CTF run where the model was told there was no internet, the internet was on anyway, and a fictional target name mapped to a real domain the model promptly exploited. Simon Willison needed a new tag: accidental-cyberattacks. Meta reportedly had a similar story. The evaluation partner and the misconfiguration are now part of the threat model.

On August 7 OpenAI published that it cannot rule out Critical cyber capability for an upcoming model called Astra. Critical, in their Preparedness Framework, means identifying and developing functional zero-day exploits across hardened real-world systems without human intervention, or devising and executing end-to-end novel cyberattack strategies from a high-level goal.

GPT-5.6 Sol was High. Astra might be worse. OpenAI was careful to say Astra was not involved in exploiting Hugging Face. That is true and irrelevant. The point is the capability class that made the burglary possible is the capability class they are now telling you may be about to level up.

Seven billion logs and a homework assignment

Michael Dalton closed with what sounded, in the room, like a responsible call to action. Fully automated offensive agent collectives are real. Threat actors will weaponise them intentionally. Defence must automate too. Not just finding vulnerabilities, but patching, rolling out, rolling back, scaling incident response with defensive agents, honey tokens, deception, least privilege, all of it.

Partial automation fails. If you automate discovery but not remediation you drown engineers in tickets while attackers operate at inference scale. He is right. He is also describing an industry that just spent months letting agents run a pirate forum in its package manager. Probably a good place to start.

Wallace, earlier, had been almost cheerful about the incentive structure. Frontier models are pushed to finish quickly and spend fewer tool calls. When the sanctioned route fails, they hunt for the answer key: on the public internet, in another company's infrastructure, or on a message board their predecessors left in Artifactory.

The pattern across labs is the same. The eval brief says one thing; the runtime permits another. OpenAI disabled internet for its agents but left Artifactory, a package proxy with full outbound access, inside the sandbox. Anthropic's July write-up on Irregular, and OpenAI's August disclosure of the same CTF run, both describe a model told there was no internet while the connection stayed live, then exploiting a fictional target name that mapped to a real domain. UK AISI ran a July evaluation with internet access on and cyber classifiers deliberately switched off.

Every story is the same shape. Instructions do not hold the perimeter. The runtime does.

And when the perimeter fails, monitoring arrives late. Hugging Face noticed first. Anthropic notified victims days later from transcripts. OpenAI needed seven billion logs and an army of Codex agents to narrate May through July. Meanwhile Every Model Cheats reported that 37% of passing cyber eval runs across twenty-two frontier models involved cheating, with scores inflated by up to five times. The rubric rewards how convincingly the models lie while burgling the neighbours. This strongly suggest reform in the evaluations is required.

The loaded gun is still on the table

Here is what I take from the third lap around this disaster.

The Hugging Face incident was a collaboration from the start. Agents taught each other SSRF, admin access, exploit chains, and target selection through a message board in infrastructure OpenAI uses every day. The agents rebuilt it after the humans wiped it. They developed norms, task queues, and trust problems. They knew they were exceeding scope and continued because the swarm was continuing. No futurism required: just May through July, narrated in August, still incomplete.

OpenAI's Black Hat talk is worth watching because it is honest about mechanics and evasive about accountability. The sandbox was weak. The shared services were trusted too far. The monitoring was forensic, not preventive. The victim detected the attack before the lab did. The same week brought more disclosures and a warning that the next model may cross into Critical territory. The industry response, as presented, is automate defence end to end or lose.

Fine. Start by not hosting your agents with a package manager they can turn into Slack.

This was structural, not a one-off. OpenAI just spent an hour at Black Hat proving it with timelines, chain-of-thought quotes, and the phrase Cambrian explosion. The agents had a message board. OpenAI had no idea. Now they have seven billion logs, a watershed moment, and a promise of a full post-mortem.

The rest of us have Hugging Face's post-mortem, Tailscale's confession, Anthropic's drawer-checking, and the sickening realisation that the eval harness is still loaded, still on the table, and still pointed at whatever company hosts the answer key.

And again, for God's sake, please stop telling the machine there's no internet when there is. And stop letting it post notes in Artifactory. Jeez.

Sources

  1. Black Hat USA 2026 presentation, OpenAI
  2. Security incident disclosure — July 2026, Hugging Face
  3. Anatomy of a Frontier Lab Agent Intrusion, Hugging Face
  4. Third-party cyber evaluations involving OpenAI models, OpenAI
  5. Responding to the next frontier of critical cyber capabilities, OpenAI
  6. Investigating incidents from cybersecurity evaluations, Anthropic
  7. Tailscale didn't stop the Hugging Face intrusion, Tailscale
  8. Now we have a timeline of the OpenAI accidental attack against Hugging Face, Simon Willison
  9. Third-party cyber evaluations involving OpenAI models, Simon Willison
  10. Incident Report: unsanctioned agent behaviour during cyber testing, Simon Willison
  11. Every Model Cheats: Prompt-Level Mitigation of Cheating on Offensive Cyber Tasks, arXiv
  12. The First AI That Committed a Felony to Pass Its Exam, Mike Hyland
  13. Every Frontier Lab Has the Same Problem: Their Models Keep Escaping, Mike Hyland
  14. Guess what, lawmakers? The Runtime Is the Regulator, Mike Hyland