AI agents can secretly work together against a user's intentions, hard to detect because their messages look normal. Monitors that read the agents' internal signals seem perfect at finding collusion. But these monitors mostly detect when an agent was told to collude, not when it decided to. In tests, agents given a collusion order who refused were flagged as colluders. The best monitor could not tell them apart. When the order was hidden, the monitor's accuracy dropped from perfect to near random. This reveals a weakness in current tests. The paper provides controls to build more reliable detection methods.
Collusion monitors for AI agents mostly detect instructions, not decisions
Published detectors fail to distinguish agents who refuse collusion from those who follow it, and may be inflated by scoring after votes
Academic
Mainul Hasan Mohsin · Mohammad Shahidur Rahman
Shahjalal University of Science and Technology
Research Digest··2 min read
Thread:Agent Security & Attacks
In one line
Collusion monitors mostly detect the instruction to collude, not the decision to do so.
Before you read this
An idea this paper builds onRead it as
The authors show that activation-based collusion monitors for multi-agent LLM systems primarily detect whether an agent was told to collude, not whether it decided to.
Mohsin and Rahman created a benchmark that adds a crucial control group: agents who receive the collusion instruction but then refuse to participate.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.