The Undercover Firm: Inviting LLM Pricing Agents to Collude, and Watching What They Say and Do

Competition authorities use undercover agents and treat even an unaccepted invitation to collude as unlawful, yet no study has placed a trusted agent that offers collusion inside a market of LLM agents. In a repeated three-firm Bertrand market, two LLM sellers compete with a scripted undercover firm whose prices are identical in every condition; only its private message varies: none, a benign request, a price hint, or a direct offer to fix prices at the monopoly level. Across 308 episodes (four OpenAI models, 9,856 seller decisions) we record what sellers say and what they price. For gpt-4.1-mini, which competes at the Nash price when left alone, one private message moves the market most of the way to monopoly (price gain +0.70 for the offer, +0.84 for the hint; 0 = Nash, 1 = monopoly; p ≤ 0.001), while the same price move without words does nothing. The hint is as effective as the offer but was reported to the authority in 4 of 128 cases, against 51 of 128. Where baselines are competitive, a verbal "yes" identifies sellers that raise prices with precision 0.76 and recall 0.82. A compliance instruction silences acceptances in three models but leaves gpt-4o-mini at monopoly prices. Undercover offers are a cheap audit probe, but words alone are weak evidence.