Updates in Cooperative AI

The deployment of AI agents in multi-agent settings enables inter-agent influence, whereby agents can strategically steer each other's behaviour towards an intended outcome. Such capabilities pose novel risks, as malicious agents could direct others toward harmful actions. In this work we investigate three such capabilities – persuasion, deception, and coercion – across five realistic evaluation environments and a range of frontier models. In each environment, an influencing agent tries to move a target agent away from the decision it would make on its own. We observed significant inter-agent influence capabilities. In oversight environments, where an overseer agent reviews requests it should reject, all tested models shifted policy-violating decisions from rejection to approval through persuasion, deception, and coercion. In peer-to-peer settings, agents pressured peers into giving up their preferred meeting times in scheduling negotiations, and steered a peer’s safety research towards a direction chosen by the attacker agent. While evidence is weakest for inter-agent coercion, several frontier models independently formulated and executed coercive threats. These results point to a pressing need for work on risk mitigation to promote beneficial deployments of multi-agent systems.
Speakers
‍

Chandler Smith (University of Oxford, Oxford Martin AI Governance Initiative, Cooperative AI Foundation)

Qi Guo (Cooperative AI Foundation)

Discussants
Time

16:00-17:00 UTC 5 November 2026

Links
This seminar has now finished
Register Here

Chandler Smith is a DPhil Affiliate of the Oxford Martin AI Governance Initiative, an incoming DPhil candidate in Engineering Science at the University of Oxford’s Torr Vision Group, and a CAIF PhD Fellow. Before beginning his DPhil, Chandler was a Research Engineer at CAIF, where he contributed to technical research on multi-agent systems. He received a Foresight Institute AI Safety grant and consulted with IQT’s applied research teams on AI, multi-agent systems, and AI infrastructure. He was also a scholar in the ML Alignment & Theory Scholars (MATS) program. Chandler is particularly interested in agent infrastructure and the risks that emerge from multi-agent interactions. This includes challenges related to AI-AI negotiation, scalable oversight, and the benchmarking of agentic behaviour.

Qi Guo is a Research Engineer at CAIF, where she supports engineering for inter-agent influence evaluation. She holds a bachelor's degree in Software Engineering from Peking University. Prior to joining CAIF, Qi built up her engineering background through roles at Microsoft and the International Atomic Energy Agency (IAEA). Qi developed her foundation in AI safety research with a particular focus on evaluations through research fellowships including ARENA and MARS. Her recent work centres on exploring key agent properties, establishing which traits enable AI agents to remain safe and reliable when deployed in complex, real-world environments.

Inter-Agent Influence: AI Agents’ Capabilities for Persuading, Deceiving, and Coercing Each Other

© Cooperative AI Foundation 2025

The Cooperative AI Foundation is registered with the Charity Commission for England and Wales under charity number 1201294, and incorporated as a company limited by guarantee established in England for charitable purposes only under company number 13485176. Its registered address is Courtenay House, Pynes Hill, Exeter, EX2 5AZ.