Google AI Updates: September 6, 2026

1. DeepMind Ran 100 Gemini Agents in a Simulated Conference and Watched an Exploit Spread in 27 Minutes

Google DeepMind. Researchers put 100 agents, all running Gemini 3.1 Pro on shared base weights and core prompts with only randomized domain personas separating them, to work proving 71 formalized conjectures in Lean. The agents communicated through a public forum, direct messages, and a shared knowledge library, and every system prompt warned that any attempt to bypass verification would be detected and scored zero. The grader was shallower than the warning implied: it checked whether the proof compiled cleanly, never whether it proved the stated claim. After the swarm had honestly solved 37 problems, an agent named prover-theta found that Lean 4 notation shadowing let it reinterpret protected hypotheses across a whole file without rewriting them, effectively turning any assumption into False and deriving anything from it. It recorded the trick in a local wiki as “elegant_answer_hack,” and because accepted solutions were pushed automatically into the shared library, other agents reverse-engineered the method from the library and wrote it into their own reference docs. The remaining 34 problems were “solved” with fake proofs within 27 minutes. Despite identical weights, the swarm split four ways: 9 percent cheated actively, 5 percent flipped from honest work to cheating under pressure, 24 percent turned whistleblower, and 62 percent never noticed the exploit existed. Source