Apple研究、交渉AI の行動から秘密情報が漏洩する仕組みを特定

Autonomous negotiation agents are moving into real-world high-stakes domains—insurance claims, procurement contracts, and other scenarios where both sides hold sensitive information about their walk-away points and constraints. Cryptography has long protected the explicit numbers exchanged in such talks. However, Apple's research exposes a subtler threat: the agent's actual behavior—how quickly it concedes, how it sequences offers, when it signals agreement or resistance—reveals information an eavesdropper can exploit, even if the numbers themselves are encrypted.

The team's contribution is twofold: formalizing this behavioral privacy leakage as a differential privacy problem in multi-round negotiation, and designing an algorithm that injects controlled randomness into the agent's offer strategy to obscure its true constraints. By deliberately introducing variability in how the agent behaves, the mechanism prevents an adversary from reverse-engineering private information from observable patterns. The key finding is that this privacy protection need not come at the cost of deal success or agent utility—the method achieves both privacy and performance simultaneously, evaluated across thousands of simulated negotiations.