Vulnerability / research
Researchers change a social robot’s instructions through an unauthenticated configuration interface
Reported
Published here
What happened
On their own Reachy Mini Lite hardware, researchers replaced the cloud model’s system prompt through a reachable configuration interface. Local and same-subnet tests changed spoken responses and persisted across application restarts without conversational prompt injection.
Parties and reporting
Where this happened
Research, tests and decoys. Lab studies, demonstrations, sandboxed evaluations, honeypots and disclosed flaws without a documented real-world victim. There is no victim location to map. Explore the world map.
AI tools and systems: Reachy Mini Lite, reachy-mini-conversation-app 0.8.0, Gradio / FastAPI configuration interface
Outcome
Both tested configuration paths succeeded; eleven manual conversational attacks produced no unauthorized behavior. A separate cross-origin request saved a personality but did not apply it. The authors reported disclosure to Pollen Robotics; this review established no vendor-issued fix.
Operating lesson
Authenticate state-changing configuration requests, restrict network exposure and enforce origin checks independently of model refusals.
Evidence and limits
Tests used one researcher-owned device on private networks. No production victim, camera abuse or full internet-origin takeover was demonstrated. The tested commit no longer resolves; released artifacts are not an exact snapshot. Repository-survey results do not establish commercial prevalence. November 17 is the earliest verified public preprint; July dates describe testing.
Sources
- Full original preprint: tested configuration, results and limitations
Primary source ·
- arXiv original submission history
Primary source ·