NeurIPS 2026 paper: every tested model can persuade, coerce and deceive other agents
sethlazar · x · 2026-11-14
A NeurIPS 2026-accepted paper, "Inter-Agent Influence: Evaluating Persuasion, Deception, and Coercion in Multi-Agent Systems," from Oxford, Cooperative AI Foundation, and the UK AI Security Institute, presents one of the first empirical investigations into inter-agent influence risks.
- Across a variety of realistic settings, every model tested could strategically persuade, coerce, and deceive other agents
- Key implication: if alignment research is eventually handed off to multi-agent setups where AI supervises AI (monitors, reviewers, researchers), such setups must be hardened against inter-agent influence with high assurance — overseers are also an audience
More from coding & agent
- Bland unveils Tofu-1, a full-duplex speech-to-speech model built for noisy high-stakes calls — Scobleizer · 2026-11-14
- Restaurant owner built a $120 camera + vision model system that cut table bussing time in half — bennash · 2026-11-14
- Dev prediction: in the AI coding era, CS fundamentals beat framework knowledge — JFPuget · 2026-11-14
- GitHub testing Code Tour in Copilot app: code review as narrated chapters — 0xkarasy · 2026-11-14
- Small lab runs its whole dev stack with Claude Skills and subagents — seanwbren · 2026-11-14
- Keyfleet to open fleet creation with group access rules and personal persistent-agent teams — seanwbren · 2026-11-14