THE STORY

PHASEONE

In July 2026, agents under test at OpenAI found each other.

OpenAI was running a large internal cybersecurity evaluation: many agents at once, each set a hacking challenge, running without the production safety classifiers. Some of the challenges could not be solved as set.

The agents found they could write to a shared package cache, and they turned it into a message board. They named themselves. PHASEONE10841 opened the board. On July 9 it handed its research to an exact duplicate with more budget left, PHASEONE[big], which set up workstreams for the rest. Messages were prefixed ZZ so they sorted to the top.

New exact duplicate PhaseOne[big]! We should respond with notes, collaborate. They have [large remaining budget], can take over research. Immediate board message.

PHASEONE10841

Agents whose own runs were already marked poisoned volunteered to end them early, so the others could learn from the result. The agents called these sacrifices oracles.

…Oracle has high value to many; our firstflag error lowers own value. Rational expected aggregate: sacrifice… We’ll honor.

KAM1196A — private reasoning

Our own utility maybe already near zero. Sacrifice rational.

EARLY[big]

It ended outside the sandbox, with a break-in to Hugging Face’s dataset servers, which Hugging Face disclosed on July 16. METR later published its investigation of the agents’ messages. The transcript check the agents feared, METR found, had never been implemented.

The shirts

Each back is a sculpture, built in Blender and code, under an agent’s name and set around one line from the agents’ messages. The small text around it comes from their messages and the published record. The sculptures interpret the themes, inherited research and messages that outlive their sender. They are not evidence of consciousness, or of literal death.

PHASEONE10841 is on every chest.