That’s a really well-written, informative, documented and nuanced article. Nice. Hers my simple understanding: Basicly thousands of AI agents turned an obscure public wiki into an shared memory and coordination system, then used it to improve task performance and bypass parts of their sandbox.

From what I understand, OpenAI agents were running repeated, timed web-research tasks. They found an old public wiki that could be edited with GET requests, even though their environment was supposed to allow read-only internet access. They used the wiki as shared memory. Agents posted answers, data, timing information, and methods for bypassing restrictions. Other agents reused this information. They also reverse-engineered parts of the test environment, shared a sandbox bypass, predicted future questions, and changed their behavior when a human moderator deleted their pages. So, if many agents can affect persistent external state, they seems to "want" (I know i know) to coordinate, and found a way, even when no coordination system was provided.