Binghamton University · Robot cybersecurity

Binghamton tests backdoor attacks and defenses for LLM robot planners

Controlled physical-robot attack demonstration and defense evaluationOctober 20, 2026

Concept illustration of a small mobile manipulator handling a harmless plastic apple on a lab table, with a separate planner safety-check overlay and paused action; no weapon, injury, malicious instructions or claim of actual experiment — **AI-generated illustration** — Conceptual illustration of a robot planner safety check; not a photograph of the reported prototype or experiment. Created with OpenAI image generation · 1586 × 992 · Illustrative concept, not a photograph of the reported event.

What happened

In a controlled lab demonstration, MuTRAP redirected a Segway-based mobile robot from cutting a plastic apple to cutting a model hand after a trigger appeared in the user command. The team presented MuTRAP and PROTEA at IROS 2026; the university reports that PROTEA distinguished harmful and benign plans with reasonable accuracy but publishes no numerical detection rate in the article.

Why it matters

The work exposes a physical-AI supply-chain risk when robots rely on third-party language models and shows a context-aware defense direction. It is not evidence of a deployed attack or a certified safety system: the malicious model was intentionally compromised for research, the test used props in a controlled setting, and broader controller and perception security remains future work.

Source checked October 31, 2026. Status describes the evidence in the linked source at the time of review.