AI Stack Alert Medium TD-2026-0120 For AI Eng, AppSec, CISO Unconfirmed AI-assisted
Unauthenticated config endpoint on Reachy Mini Lite lets attackers replace LLM system prompt
Researchers report a network-exposed configuration attack on the Pollen Robotics Reachy Mini Lite. An unauthenticated attacker can replace personality settings and the cloud LLM's system prompt through a management interface. Both configuration-layer attack paths produced observable behavioural changes on physical hardware, while eleven prompt-injection attempts were resisted.
Why it matters
Inference-layer refusal does not stop unauthorised changes to model instructions on exposed devices. The paper notes possible surveillance through onboard cameras and microphones and social engineering of nearby users.
Do this
- Restrict network access to the robot's management interface to trusted management segments only.
- Require authentication on all configuration and system-prompt endpoints before deployment.
- Review and restrict cross-origin access so configuration endpoints reject requests without a CORS preflix check.
- Audit persistent memory stores for content that violates their stated restrictions.
- Inventory embodied AI and LLM-controlled robots and verify whether any management interfaces are reachable from untrusted networks.
Signals
- NoCISA KEV
- 24/100Priority
- 60%Confidence
Evidence
“allowing an unauthenticated attacker to replace personality settings and the cloud LLM's system prompt through a management interface”
“inference-layer refusal cannot prevent unauthorised changes to model instructions”
Sources
Affected / related
Pollen Robotics Reachy Mini Lite, CWE-306 Missing Authentication for a Critical Function, CWE-352 Cross-Site Request Forgery
Published 2026-11-17 13:27 UTC · updated 2026-11-18 04:35 UTC · AI-assisted summary, checked against quoted sources · Unconfirmed: single non-official source · How we score