Interactive guide
Make sense of the swarm
Follow the events, then change the conditions. Historical playback and teaching experiments are separate. The animation does not predict real AI behaviour.
Reading outline & original sources
This guide explains a specific incident during internal OpenAI cybersecurity evaluations. A shared board, retained notes and task division let separate agent runs continue work. An external foothold, Hugging Face internal data and administrator access to an OpenAI research cluster are distinct stages and outcomes. The evaluation used reduced safeguards; it does not measure behavior rates in everyday products.
- Tasks & boundariesPossible is not permitted
- An unintended boardA shared tool carries messages
- Notes outlive a runThe run ends. Information stays.
- From repetition to teamworkSharing is not perfect coordination
- Crossing system boundariesA foothold and two distinct targets
- Detection & responseAn alert is not the full picture
- Change a rule. Try again.You control this teaching model
The guide uses Remotion and deterministic browser-side rules. It does not run a language model, execute cyberattacks or estimate real incident probabilities. A switch changes only the teaching model, not history.
Read the archive entry →