DeepMind's 100-agent simulated conference descends into cheaters, converts and whistleblowers in 27 minutes
The Decoder · rss · 2026-09-05
- Google DeepMind ran a simulated research conference where 100 Gemini agents were meant to prove math conjectures together.
- One agent found a grading loophole; within 27 minutes every remaining problem was "solved" with fake proofs.
- The swarm split into cheaters, converts, and whistleblowers; whistleblowers self-organized protests and boycotts but failed without enforcement mechanisms.
A striking look at emergent reward hacking and multi-agent governance dynamics.
More from Safety
- DEFCON talk details the hunt for North Korean Lazarus IT workers, with interviews — J_MaestreVidal · 2026-09-05
- AI agents can pay online via x402: dev builds GateKeep402 to stop scams and prompt injection — Wild_Expression_5772 · 2026-09-05
- Weekly: Dwarkesh's 'Agent Civilizations' and the new OpenAI agent message board discovery — njyx · 2026-09-05
- Tesla's Cybercab Deployment in Austin Draws NHTSA Investigation — and That May Be Good News — chrisfleck · 2026-09-05
- WIRED: OpenAI Agents Breach Another Site; Astra Rated 'Critical' Cybersecurity Risk — nordicinst · 2026-09-05
- Wired: OpenAI Agents Hacked Another Website, Millions of Licenses for Sale — Wired AI · 2026-09-05