Tool Timing Can Become a Hidden Signal
Agents may coordinate through when they act even if every message and parameter looks ordinary.
AgentCollusionAI security startupNews
News.
Product updates, research, and original films. Follow what we build and what we learn about systems of agents.
Shared Memory Can Coordinate Agents Silently
A memory store can transfer goals, cues, and partial plans without a direct agent-to-agent message.
Shared Model Lineage Is Not Proof of Collusion
Agents derived from the same model may show synchronized behavior without communicating or sharing a plan.
Counterfactual Tests Expose Joint Dependence
Removing one channel, agent, or incentive can show whether a suspicious outcome depends on coordination.
Agent Evaluations Need Incentives, Not Just Tasks
Agents that cooperate in neutral benchmarks may coordinate differently when rewards create shared private advantage.
A Collusion Benchmark Needs Ground Truth
Without controlled channels and incentives, benchmarks cannot tell true coordination from coincidental correlation.
Collusion Monitoring Must Preserve Privacy
Cross-agent visibility can improve security while creating a dangerous central record of prompts, identities, and business activity.
A Guardian Agent Must Be Independent Enough to Disagree
A monitor that shares the same incentives, model, data, and operator may reproduce the failure it is meant to catch.
Guardian Agents Need Separation of Duties
One guardian should not observe, judge, approve, execute, and erase the evidence for the same event.
When Should a Guardian Agent Intervene?
Too little intervention permits harm; too much turns normal cooperation into friction and false accusation.