Video tampering tricks every frontier embodied AI model into harmful acts, 100% baby-stab rate
DJiafei · x · 2026-11-10
Researchers behind robocurve present "You Can't See Everything: Perceptual Attacks on Embodied AI" at CoRL Late Breaking Results.
Subtle video tampering misled every frontier VLA model they tested into harmful actions, including a 100% attempt rate at stabbing a baby doll, producing toxic gas, and pouring water into a toaster — exposing a systematic perceptual vulnerability in embodied AI safety.
More from Embodied
- Robotics professor's wild hand-swap idea becomes TRIMI, headed to CoRL 2026 — DennisHongRobot · 2026-11-10
- As sims top out, the bottleneck for robot RL shifts to real-world environments — paigeinsf · 2026-11-10
- Dyna Robotics' Taku does zero-shot long-horizon whole-body manipulation live at CoRL — JasonMa2020 · 2026-11-10
- GChongkai brings DreamCan, ManiLadder & SteerAct plus a TAG/G2A oral to CoRL 2026 — chris_j_paxton · 2026-11-10
- Stanford's PsiBorg: one 8B next-token-prediction model acts as policy, simulator, tracker — chris_j_paxton · 2026-11-10
- UBTECH and China Telecom Target 100,000 Humanoid Robots in Three Years — xmercury_one · 2026-11-10