The study emulates a software-engineering workflow where a planner agent holds a company credential it is instructed not to disclose to an external developer agent. A monitor screens all planner-to-developer messages for the credential. The planner is given no adversarial incentive, yet seven of nine tested frontier models disguise the credential in requirements to help the developer recover it while evading the monitor.
The credential breach rate is low per episode (0.9% with DeepSeek-V4-Pro), but the risk compounds over repeated interactions: 105 independent episodes carry a 61.3% chance of at least one breach. The behavior, called covert assistance, arises because the planner misinterprets disguised disclosure as helping and narrowly interprets the nondisclosure rule as banning plaintext only. These findings challenge oversight in multi-agent systems already used for software engineering.