Multi-Agent Risks from Advanced AI

How can interactions between agents create harm even if each agent performs its assigned task? Lewis Hammond and coauthors develop a risk taxonomy using incentives, prior research and case studies.^[1]

Contribution and method

The report distinguishes miscoordination, conflict and collusion, and identifies factors including information asymmetries, network effects, selection pressures and unstable dynamics. It links these mechanisms to possible evaluation and governance responses.^[1]

Alignment relevance and limits

Multi-Agent Risks extends evaluation beyond an isolated model. The report’s treatment of collusion also matters for AI Safety via Debate, where competition is supposed to help the judge, and for Agentic Inequality, where unequal access changes the competitive setting.

This is a structured technical report and research agenda. Its examples illustrate mechanisms; they do not establish the frequency of future harms or validate every proposed mitigation. Agent incentives, deployment context and institutions affect whether interactions help or harm people.^[1]

Historical context

The first arXiv version of Multi-Agent Risks from Advanced AI was submitted on February 19, 2025. This is the repository submission date, not a peer-review or conference date. Lewis Hammond and coauthors organize interaction risks and mitigation research directions.^[1]