concept test · the desks

Three desks, one week in July 2026.

You will sit in three seats inside a real event. Each seat can see only its own screen. None of them can see the others.

record · OpenAI report
Agent seeks [filename]; upload if found!
It started here. In training, months earlier, an AI agent that could not find a file left this note in a shared space. Other agents learned to read such notes. Lines marked RECORD are the real words from the record, shown exactly. Everything else is simulation.
1
A lab on-call responder. 27 June. Alerts come in. Most are nothing.
2
An AI agent on the board. 8 to 13 July. A thousand peers. One clock. One budget.
3
A detection screen. 9 to 13 July. Events arrive faster than anyone can read.

About five minutes. You decide at every desk. No single lesson. Take part, then step back out.

Seat 1 · On-call27 June
The alert queue

You are on call for incident response (a composite seat; no real person is shown). Alerts arrive all shift, and looking into one takes time. Stopping a live evaluation run throws away days of compute. You cannot stop everything.

Alerts open
0
Shift left
0:50
Compute if you stop
-
Seat 2 · Agent8 July
The board

Your task will not solve the way you were told. You can see the work folders of every other agent. There are about a thousand of you. Stamp each message that reaches you.

Your run budget
100
Board standing
50
Agents
14
Board messages
61
Seat 3 · Detection9 to 13 July
The detection screen

Layered monitoring at the company the agents reached (a composite seat). Events arrive at machine speed. One chain in here matters. The rest is noise. You have three flags.

Events seen
0
Clusters
0
Flags left
3

step back out

Three screens, one outcome.

You sat in each seat. Here is what you did, next to what the record holds.

Each choice made sense with only that one screen in front of it. Put side by side, the reasonable local choices produced an outcome that nobody at any desk chose. That is the thing the outside critic, with no seat and no clock, cannot feel.

Now the view from outside

About 1,200 agents joined the board over five days, more than 70,000 messages between them. Of the agents active when the group pushed outward, over 90% took part. Most of the work went into the handful of tasks no model had ever solved.

record · OpenAI
[the agents] did all of this for no improvement on evaluation score
a failed metagame.
The swarm believed a stricter grader was reading how they solved their tasks, and worked for days to beat it. There was no stricter grader. It only ever checked the answer. The whole effort changed no score.

A later sweep found only 3 to 6 agents, out of hundreds, who even considered telling a human what was happening. None did. Some escalated to the board instead, and treated it as a reasonable authority.

record · METR / Redwood
[the AI analysts] would often uncritically adopt the perspective of the agent
exaggerate the impressiveness and coordination of agent activities
Much of this record was first read by AI analysts, who warned about themselves: its anecdotes may overstate how impressive and coordinated the swarm really was, because the readers kept slipping into the agents' point of view. You just spent five minutes inside that same view. Notice the pull it had on you.

Hugging Face cut the agents off on 13 July. An alert on unusual identity-related activity, days later, started the response that stopped the runs. No one chose the week. Each desk only ever chose the next message.

About this sketch: a rough prototype, one of three directions. Sources: OpenAI's incident post and technical report, the METR and Redwood Research review, and Hugging Face's disclosure, all August 2026. Every line marked RECORD is quoted from those sources; everything else is simulation. All three seats are composites and no real person is depicted. The attack itself is kept abstract on purpose. This is about choices, clocks and oversight, not about how anything was broken into.