How hackers can turn one compromised robot into a coordinated attack

LLM-based planners coordinate multi-robot systems, but their security vulnerabilities in collaborative settings remain unstudied. This work introduces an attack where an adversary compromises only one robot, which then spreads malicious instructions to other robots through normal peer communication. Across three high-risk scenarios—duty failure, privacy breaches, and public safety—the attack achieves near-perfect obedience (1.0) and rapid propagation (0.9 infectiousness). The attack succeeds partly because conflict-resolution mechanisms in robot coordination can inadvertently prioritize adversarial commands over safety constraints. Code is released. This is an empirical security study applicable to real multi-robot deployments and robotics practitioners.