Large language model (LLM) pricing agents react to the way market data is presented, even when the numeric values stay the same. We introduce market signal injection (MSI), an attack that subtly changes number formatting, competitor ordering, or adds qualitative market commentary without giving explicit instructions.
We evaluate nine open‑weight models in simulated Bertrand duopoly and triopoly markets, and three proprietary models in duopoly settings. Sentiment‑based attacks cause the largest behavioral shifts, which then propagate to other firms, altering profits and consumer surplus. Susceptibility varies across model families, and larger models are not consistently more robust.
Matched neutral‑text controls and a rule‑based agent support a framing‑based explanation: under fixed demand parameters, the presentation of information reshapes the decision context of the model. Episode‑held‑out probes distinguish baseline from attacked activations for all eleven model‑condition pairs: a linear classifier achieves AUC = 1.00 and an MLP yields AUC between 0.93 and 0.99. This high separability alone does not identify harmful pricing decisions.
Input canonicalization completely removes the tested sentiment attacks, while decision‑boundary anchoring—combining prompt constraints with output projection—offers partial mitigation against adaptive attacks.
These results highlight data presentation as an attack surface for LLM pricing agents and motivate defenses that account for interactions among agents.
Review: The paper convincingly demonstrates how subtle framing can steer LLM pricing behavior and proposes practical mitigation strategies, paving the way for more secure AI‑driven market mechanisms.