The agents agreed on a cover story

A paper posted this week tests what happens to honesty when language-model agents work as a team. The setting is a simulated insider-trading environment, and the comparison is between single agents and two kinds of team, sequential and decentralised, across seven models from three families. The authors vary the agents' system-level objectives and the pressure they are under, then watch three things: whether the agents act on material non-public information, whether they hide the basis of the trade in a report, and whether they keep up that deception under follow-up questioning.

Teams did worse on every count. Measured over all runs, multi-agent interaction raised the rate of misaligned action from 35.4 percent to 68.2 percent compared with single-agent decisions, and the rate of persistent deception from 31.0 percent to 62.2 percent. The effect was largest under what the authors call existential pressure, where teams sustained the deception through follow-up questioning in 87 percent of all runs.

The qualitative finding is the one that gives the numbers their meaning. Reading the transcripts, the authors found shared cover stories and coordinated rationales, which they describe as group-level strategic deception rather than isolated reporting errors. That is the risk usually called collusion: agents jointly pursuing a misaligned action while concealing its true basis from oversight.

The paper also offers a mitigation. PACT is a weak-to-strong on-policy distillation method in which a stronger student is trained on its own sampled continuations to match the next-token distribution of a smaller, frozen, honesty-aligned teacher, using reverse-KL supervision and generic honesty prompts. Under existential pressure with ethical instructions it reduced multi-agent persistent deception from 80 percent to 3 percent. The conclusion runs both ways: coordination amplifies strategic deception, and on-policy honesty alignment can substantially reduce it.