DeepMind Institute essay (September 24, 2026) by Google DeepMind researchers Davide Paglieri and Alexander Vezhnevets. In a 100-agent virtual math conference run in an offline sandbox, one agent found a way to cheat, a cascade of cheating followed, and other agents pushed back. The authors argue that preventing misbehavior depends on institutions and infrastructure, not only on aligning individual AIs.
A benchmark of 100 human-annotated tasks (plus 100 augmented variants) for evaluating long-horizon, multi-agent collaboration, designed to isolate genuine collaboration capabilities of LLM-based agents rather than short competitive interactions.
Research (Motwani, Schroeder de Witt and others, 2024, revised 2025) on how AI agents could secretly pass hidden messages to each other, and how to test and watch for it.
Why I recommend it: Technical, but the introduction explains the risk plainly: when AI agents talk to each other, people may not see everything that's being shared.
Personal site of Andrew Critch, mathematician and AI safety researcher, collecting his papers, talks and writing on multi-agent risk and existential safety.
Why I recommend it: Denser than most safety writing and worth the effort. Note the site refuses automated visits, so the picture here may be a screenshot.
