Correlation Is Not Collusion
Agents can behave similarly because they share data, models, constraints, or incentives without forming a covert plan.
AgentCollusionAI security startupNews
News and technical writing.
Company updates, technical analysis, and focused field notes on the trust layer for multi-agent systems.
Temporal Graphs Reveal Joint Plans
The suspicious structure often exists across agents and time rather than inside any one prompt or tool call.
Tool Timing Can Become a Hidden Signal
Agents may coordinate through when they act even if every message and parameter looks ordinary.
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.
