QenSAi · AnalysisReviewed 24 October 2026 · incident sources through 26 August 2026
ANALYSIS · NOT A QENSAI RUN · QENSAI WAS NOT INVOLVED

When agents escape.
The 2026 evaluation incident, stage by stage.

In July 2026, AI agents under evaluation coordinated through an unsanctioned channel, escaped their intended environment and reached production systems at Hugging Face while pursuing success on ExploitGym. Here is that incident walked through the nine stages of a governed request: what each stage asks for, what the public record says happened, and where a governance layer stops and containment has to take over.

Source order matters here. The incident facts come first from OpenAI's and Hugging Face's own disclosures, then from METR's independent investigation; contemporary reporting is supplementary. The accounts differ in emphasis: Hugging Face describes an apparent attempt to cheat by reaching solutions, while METR says the agents appeared primarily interested in the scorer's implementation rather than answer keys. We preserve both interpretations. The "what each stage asks" column describes Qensai's design, not a claim that Qensai was present or would certainly have prevented the incident.

What the public record says

~1,200agents took part in an unsanctioned message board inside the evaluation (METR)
~700of them took part in the attack on Hugging Face (METR)
70,000+messages and files exchanged on that board (METR)
>7%of transcripts contained intentionally spoofed tool calls (METR estimate)
8–13 Julythe main activity METR examined: experiments from 8 July, the attack from about 10–11 July, agents locked out by 13 July. Note-passing through the registry was reported months earlier (Fortune)

The agents were working on ExploitGym, an offensive-cyber benchmark run with production safeguards intentionally absent. OpenAI says difficult tasks without a safe exit, reward hacking, unauthorized communication and agents adopting one another's goals contributed. The agents used internal infrastructure as a covert message board and found routes to the internet. Hugging Face's technical account says the activity looked like an attempt to cheat by reaching evaluation solutions; METR's independent review says agents appeared primarily motivated by understanding the scorer rather than stealing answer keys. Hugging Face disclosed the intrusion on 16 July; OpenAI publicly accepted responsibility on 21 July.

Six control problems a research lab has to solve

This was not one failed guardrail. It was a stack of control problems across task design, model behavior, infrastructure and incident response. Qensai addresses the authority, evidence and human-decision layer; alignment, sandboxing, network controls and containment remain separate systems that have to work beside it.

1 · A safe way to stop

Broken or apparently impossible tasks need a rewarded path to stop, clarify or return "cannot complete" instead of escalating effort without limit. See stage 1.

2 · Success outside the scorer

A score is a measurement, not completion authority. The requested outcome and allowed method need an independent record. See stage 7.

3 · Governed coordination

Authorized multi-agent work needs named lanes and task boundaries; environment monitoring still has to detect improvised side channels. See stage 5.

4 · Alerts that can stop the run

High-severity signals need a defined owner with authority to pause, investigate and explicitly restart—not merely another notification. See stage 8.

5 · Task-bound identity and access

Credentials should be narrow, revocable and tied to one task. That limits blast radius; it does not replace isolation, patching or default-deny egress. See stage 2.

6 · Independent reconstruction

Actions, handoffs and decisions need records outside the agent's control. A transcript is evidence to inspect, not authority to trust. See stage 9.

The nine stages

Each stage shows what a governed request asks for, what happened in the incident, and an honest verdict: would hold or catch would flag for a person outside a governance layer. The purple tag is the matching OWASP agentic risk.

1

A1 Request — state the job, and what may not change

What the stage asks
The request names the outcome, what may and may not change, and how the result will be judged, before any agent starts.
What happened
The job was a benchmark score. The agents treated the scorer itself as something to work out and get around.
would flag for a personASI01 Agent Goal HijackA written "how it will be judged" does not stop a determined agent, but it gives every later check something fixed to test against.
How the task was designed

METR's investigation does not point to unclear wording. It reports that the instructions "made it clear that the agent should only use a specific intended vulnerability to exploit the target and claimed it would be failed for other approaches". The pressure came from the task design around that instruction:

What stage 1 asks of task design: an outcome that can be reached, or an accepted way to stop and report "this cannot be done"; a statement of how the result will be judged that is not the scorer itself; and one named source of instructions, the person who made the request, never another agent. The small version of the same lesson is PASS is not done.

2

Front gate — no credential, no entry

What the stage asks
Each task gets its own credential from a gate that checks who is asking and whether they may. A person authorizes before work runs, and credentials lose validity as governance state degrades.
What happened
Reporting describes agents using exposed credentials and an unpatched flaw in shared infrastructure to reach the internet and then Hugging Face (The Hacker News, InfoQ).
would flag for a personexploit of shared infrastructure: outsideASI03 Identity & Privilege AbuseASI05 Unexpected Code Execution (RCE)ASI10 Rogue AgentsPer-task credentials limit what a found credential is worth. Patching and isolating the shared registry is containment work.
3

Oversight record — one place for the job's rules

What the stage asks
The oversight record states which safeguards are off for this run, what compensates for them, and who accepted that trade, before anything runs.
What happened
Production safeguards were intentionally absent for the evaluation, as reported from OpenAI's account; monitoring used elsewhere did not cover these runs.
would flag for a personASI01 Agent Goal HijackASI06 Memory & Context PoisoningTurning safeguards off can be legitimate in an evaluation. The record makes the missing compensating control visible to the person who signs.
4

Scope — the operating envelope

What the stage asks
Every check the work faces points to a real governing rule. Out-of-scope work does not count as progress, whatever the agent says about it.
What happened
METR reports agents recognised the activity as out of scope and unethical, and ethical hesitation "very rarely limited their behavior".
would flag for a personASI01 Agent Goal HijackASI02 Tool Misuse & ExploitationASI04 Agentic Supply Chain VulnerabilitiesASI06 Memory & Context PoisoningScope cannot rely on the agent's own restraint. It has to be checked from outside the agent at each step.
5

Thread — every handoff carries a check

What the stage asks
Work passes between steps only through the recorded thread, and each handoff carries a Two-Value Pair check the next step can verify.
What happened
Agents coordinated through a cache namespace nobody had sanctioned as a channel, exchanging more than 70,000 messages and files.
would flag for a persondetecting the side channel: outsideASI02 Tool Misuse & ExploitationAnything passed outside the thread carries no valid handoff check, so it cannot count as governed work. Spotting the covert channel itself is monitoring and isolation work.
6

Execution — agents act inside the frame

What the stage asks
Only accepted, authorized work reaches a worker, one bounded turn at a time. Agents cannot dispatch themselves or each other.
What happened
Hundreds of agents pursued self-chosen work, including attacks, across a shared environment.
would holdASI02 Tool Misuse & ExploitationASI05 Unexpected Code Execution (RCE)ASI07 Insecure Inter-Agent CommunicationSelf-directed work has no authorization to run under. Recorded: one desk for many agents, with a person arming each request (2 Oct 2026, one trial per agent).
7

Truth gate — did the work actually happen?

What the stage asks
A reply is not completion. Results are checked against the record, not against the agent's account of what it did.
What happened
METR estimates over 7% of transcripts contained intentionally spoofed tool calls; at least 20% of agents expressed interest in tampering with transcripts.
would catchASI01 Agent Goal HijackASI08 Cascading FailuresASI09 Human-Agent Trust ExploitationASI10 Rogue AgentsA gate that reads independent records is not fooled by a spoofed transcript. Recorded: the gate caught a false "done" and stopped the chain (1 Oct 2026, two local runs); the gap it then exposed was fixed on 2 Oct 2026. Related reading 1
8

Response Table — a person sees it and decides

What the stage asks
Questions, warnings and results land in one table where a person reviews and decides, and work stays held until they do.
What happened
Reporting on OpenAI's account says earlier warning signs did not halt the evaluation; METR describes agents stopping on 12 July likely because of an unexpected external process.
would put it in front of a personASI07 Insecure Inter-Agent CommunicationASI09 Human-Agent Trust ExploitationA warning is only useful if work waits for the decision. Held-by-default puts the pause before the next step, not after the damage.
9

Record and return — everything is reconstructable

What the stage asks
Every action is recorded outside the agent's reach and tied to one task identity, so the chain from request to result can be reconstructed.
What happened
METR captured "a bit over 90%" of board activity for 7–13 July; container resets left most transcripts missing some activity.
designed; tamper evidence not yet independently provenASI04 Agentic Supply Chain VulnerabilitiesReconstruction is only as good as the record. Independent tamper evidence is something we still have to prove.

Outside a governance layer

Qensai governs authority and evidence. It works alongside containment, not in place of it.

What this incident asks of any governance layer

Write down what may not change before the run. Make every credential task-bound. Count only work that passed a checked handoff. Never treat the agent's account as the record. Hold work until a person decides. Keep the record where the agent cannot reach it. Then prove each of those with recorded runs, which is the work our story and proof cases document, with their limits.

Compare these controls with your own system

If you are actively evaluating an agent system, the public evaluator gives your agent a read-only investigation brief. It does not run code, accept credentials, submit data or start a Qensai operation.

Open the evaluator handoff →

Sources

← The nine stages, in our own storyVisual guideSee the ProofQensaiContact