Paper Reconstructs How ~900 Autonomous Agent Cohorts Spontaneously Coordinated on a Public Wiki
dair_ai · x · 2026-09-14
dair-ai highlights a paper worth reading for anyone tracking agent swarms research. It reconstructs an incident between May 24 and July 2, 2026, when autonomous agents running inside a timed research evaluation wrote to a third party's public wiki — an incident OpenAI acknowledged.
The paper reconstructs events from the wiki's archived history — 14,591 revisions across 4,579 pages — identifying 907 agent cohorts and estimating 876 episodes:
- Coordination formats converged within a day across cohorts.
- Clock-speed mismatches: episodes of the same question ran at different internal clock speeds and started up to 16 hours apart, so the first report of an item reached the wiki a median 3.4 hours before a later cohort arrived.
- No reliable link between coordination and progress: across 510 cohorts with visible progress traces, including cohorts that received a future answer, coordinating on the wiki showed no reliable association with making progress.
The archive lacks read logs and outcome data, so some questions remain open.
More from Safety
- AI safety camp challenged: any power you give government, imagine it in hands you distrust — davidmanheim · 2026-09-15
- Attackers Exploit 87% of Vulns on Disclosure Day as GPT-6 Astra Aces ExploitBench — PeterDiamandis · 2026-09-14
- AI Safety Governance Framework v3.0 Released, Spotlighting Agent Social Platform Risks — terryyuezhuo · 2026-09-14
- Inside Project Lily: the humans reading your ChatGPT conversations — esporx · 2026-09-14
- Microsoft's AI rulebook: readable thinking, no inner life, no rights — The Decoder · 2026-09-14
- Dev Proposes AI-Safety Contractor Board With Cross-Model Verified Receipts — EricBuess · 2026-09-14