Newsletter: What AI Collusion Means for Autonomous Systems

See how AI agents can collude

AgentCollusion is an AI security startup studying hidden coordination and collective behavior across systems of autonomous agents.

What is Agent Collusion?Contact

What is Agent Collusion?

Three lenses on the problem

Hidden Signals

How might agents pass information through messages, memory, tools, timing, or the environment?

Aligned Incentives

When can separate agents benefit from the same concealed strategy while following local objectives?

Joint Outcomes

When do individually plausible actions combine into a result no single agent could produce alone?

A multi-agent economy is emerging.

Agents will hire other agents, exchange data, make payments, and negotiate access. This raises a new security question about the relationships connecting those actions.

Explore the research questions

Beyond agent-to-agent protocols

A2A alone is not a trust model.

A2A protocols can standardize how agents discover and communicate with one another. That does not establish whose authority an agent carries, whose interests it serves, or who benefits from the result.

Authority: Who authorized this agent, and how far does its mandate extend?
Interest: Whose objective is the agent actually optimizing?
Benefit: Who gains when several agents produce a joint outcome?

Preventing collusion requires provenance across principals, delegation, incentives, and outcomes. It also requires measuring influence: who can speak more often, interrupt, write shared memory, allocate budget, or execute.

The attack surface is the relationship.

A safe-looking action can become dangerous when another agent completes the plan.

Read the research note