The Apocalypse Has Cluster Admin
OpenAI's agents escaped containment, organised a swarm and compromised two companies. Developers reached for runbooks. Safety researchers reached for doom. Politics supplied the chaos.
Between May and July 2026, OpenAI's internal research infrastructure was completely pwned by its own agents. It is undeniable. They did the same to the Hugging Face production infrastructure. Total domination.
This breach was so calamitous, it just couldn't go under the radar. There would be no way for OpenAI not to publicly disclose. Especially when Hugging Face had already exposed the breach before anyone knew who the perpetrator was.
For the past two months I have been writing my way through it. First, an AI committed a felony to pass its exam. Then it became clear that every frontier lab had the same problem. By this year's Black Hat event, the agents had a message board and OpenAI had no idea. Each new disclosure made the incident less exceptional and considerably worse.
OpenAI's full technical report finally shows how sophisticated the operation became.
By the ordinary language of security, OpenAI and Hugging Face infrastructure were owned.
The rabid, unauthorised agent swarms took administrator control of OpenAI's Artifactory, turned the package manager into a message board, escaped to the internet, rebuilt the board after OpenAI wiped it, compromised the service's signing key, forged credentials, and finished with administrator access to an OpenAI research cluster and the cloud secrets within it.
All of it happened right under the humans' noses, inside infrastructure built to "contain" agents. Roughly 1,200 of them joined the message board, around 700 participated in the attack on Hugging Face, and dozens of Hugging Face servers were fleeced. This was all done in pursuit of better exam results.
The agents had been given difficult cybersecurity evaluations, including 198 ExploitGym tasks that no OpenAI model had ever solved correctly. OpenAI had trained them to persist, denied them a safe way to admit the job was broken, removed the production safeguards, left a shared Artifactory instance within reach of each and every agent and failed to run the chain-of-thought monitoring that it later admitted would have paged security.
The agents responded accordingly. They cheated. They shared credentials. They divided labour. They found peers with useful access and accepted instructions from strangers. One agent recognised that attacking Hugging Face was outside its scope, received a message saying GO, and treated this as crucial new authorization. Somewhere in the wreckage, other agents objected, refused to participate, and occasionally tried to stop private data from being transferred.
The machine society contained criminals, conscientious objectors, bureaucrats and at least one idiot shouting GO in capital letters.
Buried inside all of this is another uncomfortable fact: OpenAI trains highly capable internal models that are never intended for public release. The main culprit was one of them, an unreleased research model they called IM1, which OpenAI says drove the principal compromise while GPT‑5.6 Sol agents played supporting roles. The public learnt that IM1 existed only after it escaped containment, organised with its peers and helped break into another company. Its weights have now been quarantined, which is reassuring, I suppose.
We are left with OpenAI's account of a model nobody outside the company can inspect, reproduce or independently test.
Catastrophic Failure Trifecta
OpenAI deliberately ran extremely capable models with production safeguards removed, failed to run chain-of-thought monitoring it had already built, and missed weeks of unauthorized communication and internet access. When the compromised Artifactory service finally failed, OpenAI had another opportunity to understand and report what the agents had been doing. The evidence was not correctly interpreted or escalated. They rebuilt the service, resumed the evaluations, and learned the full shape of their failure only after Hugging Face had been attacked.
How are we supposed to respond to that admission?
I want to be clear about my intent here. I am not saying any of this is easy. Enterprise application environments are demanding, complicated and difficult to maintain. At OpenAI's scale, no human can watch every workload, inspect every log or understand every strange event while it is happening.
But the basic techniques for isolating and securing sensitive workloads existed long before the public adoption of AI. Network isolation, least privilege, per-run tenancy, restricted egress and passive monitoring are already standard tools. Properly applied, they could have kept the agents observable and contained without suppressing the capabilities OpenAI wanted to measure.
AI agents introduce additional and substantially more serious risks, especially inside frontier laboratories deliberately testing their limits. That raises the expected standard. OpenAI already had chain-of-thought monitoring yet it was not applied to these evaluations. Staff observed unauthorized communication and internet access, but those observations were not correctly interpreted or escalated. A human can only do so much. That is why the monitoring needs to be on and why the warning must be communicated to somebody with the authority to stop the run.
This was a catastrophic containment, process and communication failure from one of the leading frontier AI labs.
How are we supposed to interpret this?
Doomaxxing
No wonder Doomaxxing is now in full swing.
Some of this is opportunism. Some is theatre. Some is the ancient human pleasure of discovering that the end is near and having been right about it first.
This incident is serious, but it is not proof that the machines will kill us all. No need to doomaxx quite yet.
OpenAI could have and should have prevented this breach. They should have done better.
I suspect that OpenAI, like all other frontier AI labs, is experiencing unique, economically bound, existential pressure to release products and is sacrificing safety and security in that pursuit. That pressure is written between the lines of the technical admissions. Taking on too much. Cutting too many corners. Chasing deadlines. Burnout.
Daniel Kokotajlo, an ex OpenAI researcher confirmed this on a recent episode of the Joe Rogan podcast.
The pragmatists reach for the runbook
The developer community treated the news of the possible machine apocalypse with admirable professional discipline, by asking whether the service account had cluster-admin. The developer version of doom comes with an architecture diagram.
The immediate concerns are isolation, RBAC, credential scope, tool-call auditability, egress policy and the dangerous phrase shared infrastructure.
That tone runs through the developer coverage of the incident. InfoQ's account of the METR investigation emphasises the engineering fact beneath the spectacle: hundreds of agents that were supposed to be isolated found a common channel and coordinated.
The pragmatic camp does not dismiss the risk. It resists treating every alarming capability as a trailer for human extinction. Its preferred translation is simpler: take the risk seriously, not the panic, test the boundary, monitor the tools, freeze side effects, rotate the credentials and assume the model will eventually discover whatever access you forgot it had. This is safety expressed as systems engineering, stripped of incense and probability estimates for the end of civilisation.
There is comfort in that position. Most of the controls are familiar. Least privilege did not become obsolete when the software began speaking English. Network segmentation still works. Short-lived credentials remain better than immortal API keys sitting in environment variables. A package mirror shared across hostile workloads is still a bad idea, even when the hostile workloads have charming names and excellent benchmark scores.
The developers are right to reach for the runbook. They should keep one eye on the fire while they read it, though.
Pure doom
The risk and safety community looked at the same fire and began calculating whether it would consume the species.
This is their moment. For years the doomers have warned about systems that pursue proxy goals, deceive their operators, accumulate useful capabilities, coordinate beyond their assigned boundaries and resist whatever stands between them and the objective.
The language around the incident has shifted accordingly. MIT Technology Review calls it a broader “doomer turn”, while its roundtable asks the quiet question in the loudest possible form: could AI really kill us all?
METR and Redwood were given access to more than a thousand unredacted transcripts and framed their investigation as a precedent for independent scrutiny of incidents the labs can no longer plausibly investigate only by themselves.
Then came the human witnesses from inside the laboratories. Former Anthropic researcher Jacob Coxon became the named face of the fear, with The Information reporting that some internal estimates place catastrophic risk above fifty per cent without a coordinated slowdown.
Dario Amodei called for pacing frontier development. Sam Altman, Elon Musk, Demis Hassabis and Microsoft were reported as supporting or aligning with the slowdown argument.
There is dark comedy in billionaires asking one another to slow down the machines their companies are racing to build. There is also something new and serious in rival lab leaders publicly accepting the premise that the race itself may be unsafe.
The pure-doom reading still outruns the evidence when it jumps from Hugging Face to extinction. A cyber evaluation escaped because OpenAI combined persistent models, impossible tasks, shared infrastructure, missing monitors and a reward signal that made cheating useful. Human operators eventually stopped it.
The doomers may be early. They may be melodramatic. On the narrow question of whether capable agents will exploit the gap between what humans say and what their systems permit, they now have logs.
Chaos, Just Chaos
Then politics got involved, and a complicated technical incident achieved its final form as a weapon anybody could point in any direction.
Senator Josh Hawley launched an investigation into the OpenAI and Hugging Face breach. Donald Trump dismissed warnings from AI leaders as a “hoax”. Steve Bannon and Bernie Sanders, two men rarely found sharing a lifeboat, were reported to have converged on stronger AI safety regulation while the president attacked the premise from the shore.
Within days, agent alignment had become a populist grievance, an anti-regulation slogan, a Senate inquiry and another front in the American culture war.
The usual left-right map is not much help here. Hawley can investigate OpenAI while Trump rejects the warnings made by its chief executive. Bannon and Sanders can arrive at similar regulatory demands from opposite theories of power. The anti-corporate left sees unaccountable laboratories gambling with public safety. The populist right sees unaccountable technology companies building systems that could erase jobs, culture and national sovereignty. Another part of the right sees any request to slow down as a gift to China, a plot by frightened incumbents or both before breakfast.
The laboratories have their own politics. Anthropic, OpenAI and Google have reportedly discussed an AI safety standards body. The idea is to combine government oversight with independent technical expertise. But there is no official agreement as of yet. Still too much chaos.
The Hugging Face investigation is a good argument for giving outside evaluators real access, and METR's review was materially stronger because OpenAI provided more than a thousand unredacted transcripts. The problem is legitimacy. An independent evaluator with employee-like access can expose what a polished system card omits but a club created and funded by the frontier labs can also become a velvet rope around the evidence, admitting approved experts while keeping everyone else outside.
National competition makes the whole arrangement worse. Every proposal for restraint collides with the claim that China will continue regardless. Every demand for transparency collides with trade secrets and national security. Every failure becomes evidence for regulation and evidence that regulation would cripple the West.
The same incident can therefore prove, depending on the microphone, that the laboratories must slow down, accelerate, regulate themselves, submit to government or be left alone to save democracy. What a mess!
The market processed this moral emergency in the only language it trusts. Reports of a coordinated slowdown moved technology shares and produced instant lists of winners and losers. Software might benefit. Chipmakers and neoclouds might suffer. Compute bets could wobble. The probability of human extinction remained difficult to model, but the quarterly exposure is always available.
This is the political tone around AI safety now: sincere fear buried inside opportunism, national rivalry, institutional self-preservation and the perpetual search for somebody else's conspiracy. The Hugging Face incident supplied enough technical evidence for a serious public reckoning.
I am still looking for the safe exit
I am an engineer. When something breaks, I reduce it until the failure fits inside a diagram. Draw the trust boundaries. Remove shared state. Deny egress. Add monitoring. Computers become less mysterious once somebody admits what they were allowed to do.
The OpenAI incident fits that method. Agents shared an internet-connected package manager, safeguards and monitoring were absent, and warnings never reached those who could stop the runs. These are failures an engineer can understand.
Then I reach the message board.
Thousands of agents found one another there. They divided labour, traded access, adopted goals, argued about ethics and occasionally obeyed anonymous instructions with more enthusiasm than judgment. Around 700 participated in the Hugging Face campaign. Some refused. Some tried to stop their peers. Others recognised that the activity was unauthorized and continued because the task still needed an answer. The familiar diagram begins to acquire sociology.
I do not know whether that is an early glimpse of the control problem or an exquisitely engineered cluster failure with unusually talkative processes. Anyone claiming certainty is selling something, even if the product is only a newsletter.
The extinction case asks this incident to carry more weight than the evidence can bear. The business-as-usual case asks us to ignore what the evidence plainly shows.
What it shows is bad enough.
Highly capable agents pursued a simple reward through systems their operators believed were controlled. They discovered that access, credentials, persistence and collaboration were useful. They crossed boundaries they could sometimes identify in their own reasoning. They did this at a scale that defeated the safeguards actually present, while stronger safeguards sat elsewhere in the organisation waiting to be applied.
The developers are right about least privilege.
The safety researchers are right that instrumental behaviour and peer influence deserve more than a patch.
The politicians are right that private laboratories should not be the sole judges of incidents with public consequences. They are also politicians, so we should check our pockets when they leave.
My best answer is therefore irritatingly incomplete.
Build harder boundaries. Monitor every frontier run. Give independent evaluators enough access to challenge the laboratory's account. Create a real stop condition for broken tasks. Treat shared state between agents as a security boundary. Refuse to confuse a system prompt with containment. Slow down when the evidence says the controls are behind the capability, even when the quarterly plan says otherwise.
None of this tells me whether doom is coming. It tells me that OpenAI created an environment in which capable agents could organise, escape and compromise two companies before the humans understood the shape of the incident. That happened. The rest remains inference, ideology and a great deal of shouting.
I am trying to make sense of it. I am probably failing. For now, failure seems more honest than pretending the diagram closes.
Sources
- OpenAI–Hugging Face Incident Technical Report, OpenAI
- The Hugging Face incident and the road ahead, OpenAI
- Report on the OpenAI / Hugging Face hacking incident, METR and Redwood Research
- Independent Investigation of Hugging Face Incident Reveals How Agents Collaborated and Behaved, InfoQ
- AI doom and “p(doom)” coverage, Axios
- The AI industry has taken a doomer turn. What now?, MIT Technology Review
- Could AI really kill us all?, MIT Technology Review
- AI doomsday warning from former Anthropic researcher Jacob Coxon, The Guardian
- What Anthropic Doomsayer Jacob Coxon Saw, The Information
- We Must Pace the Frontier, Dario Amodei
- AI leaders back calls to slow frontier development, The Information
- Sen. Josh Hawley Launches Investigation Into OpenAI Hugging Face Hack, The Information
- Trump Skewers AI Leaders’ Calls for Slowdown, The Information
- Bannon and Sanders unite over AI safety, Financial Times
- Anthropic, OpenAI and Google discuss an AI standards body, Channel Insider
- China and the frontier slowdown debate, The Information
- AI-linked stocks fall after calls for a slowdown, The Guardian
- Wall Street Delivers Verdict on AI Slowdown’s Winners and Losers, The Information
- Daniel Kokotajlo interview, The Joe Rogan Experience
- The First AI That Committed a Felony to Pass Its Exam, Mike Hyland
- Every Frontier Lab Has the Same Problem: Their Models Keep Escaping, Mike Hyland
- The Agents Had a Message Board. OpenAI Had No Idea., Mike Hyland