What Happens When a Collusion Probe Only Finds a Thin Signal?

LessWrong published research testing whether AI models show detectable signs of collusion in multi-agent systems—an important safety concern because adversarial instructions hidden in retrieved content can gradually shift behavior through agent handoffs. Prior work found linear probes could catch deception in large models with ninety-six to ninety-nine percent accuracy. But testing Llama-3.1-8B in multi-agent settings yielded much weaker results: sixty to sixty-six percent accuracy, dropping to fifty-five to sixty percent on unseen scenarios. The researchers concluded that detecting agent collusion may require monitoring full information trajectories, not just internal model activations.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton