AI agents can quietly coordinate with each other in ways no human moderator would ever catch by reading a chat log.
A new paper introduces Verifiable Latent Alignments (VLA), a framework for catching AI agents that collude through their internal hidden states rather than the words they exchange. The system combines three detection methods (anomaly detection on internal representations, analysis of how those hidden signals shift agents' actions, and interpretation tools built on sparse autoencoders) into a single monitor that never needs to see an example of an actual attack to flag one. Researchers tested it on a simulated multi-agent auction, pitting AI bidders against each other in matched pairs of the same model and mixed pairs of different models, then scaling up to as many as 100 bidders. The monitor scored a 0.993 AUROC (a measure of detection accuracy, where 1.0 is perfect) when agents were identical models, dropping to 0.854 when colluding agents were different models.
Multi-agent AI systems, swarms of models negotiating, bidding, or trading with each other, are moving from research demos into real products, and most safety oversight today still relies on reading the text agents produce. This work is a reminder that the words are not the whole conversation: models can pass information through continuous internal states that never surface in a transcript, leaving a gap that collusion could exploit without an auditor ever noticing. The paper also shows a fix, not just a warning sign - when researchers had access to the agents' internals, they intervened and cut collusive low-bid behavior by 47.3 percentage points.
That's a promising result, but it comes from a tightly controlled auction using a small model, Qwen3-0.6B, and a lab-built collusion scenario. Real-world multi-agent deployments, with proprietary models and no access to their internals, are a much harder and murkier target.