Network Exposure Configuration Attacks Against LLM-Controlled Social Robots
Network-Exposed Configuration Attacks on LLM-Controlled Social Robots
Researchers demonstrate a network-exposed configuration attack on the Pollen Robotics Reachy Mini Lite robot: an unauthenticated attacker, without adversarial conversational input, can replace personality settings and the cloud LLM's system prompt through the management interface and cause observable behavior changes on the physical hardware. The paper reports that 11 prompt injection attempts across five categories were all resisted, while both configuration-layer attack paths succeeded. The attack supports local and network-adjacent access, corresponds to CWE-306 and CWE-1327, has a CVSS v3.1 score of 6.5 (medium), and the configuration endpoint also has a cross-origin issue with no CORS preflight. A review of 59 public repositories found that 10 of 11 real applications with verifiable source code lacked authentication, or 11 when framework defaults are counted; the authors say this indicates the problem recurs in the sampled repositories rather than being prevalent in commercial products. The paper notes that inference-layer refusals cannot prevent unauthorized changes to model instructions, and potential consequences include surveillance through onboard cameras and microphones and social engineering of nearby users.