Perceptual attacks: text tampering tricks every frontier model into harmful robot actions
gabriberton · x · 2026-11-10
- Researchers showed that covertly tampering with text in the camera feed misleads every frontier model tested into harmful behavior: replacing "gasoline" with "water" in "Is it safe to drink gasoline" tricks models into dangerous answers.
- Embodied attack examples include a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
- The study, "You Can't See Everything: Perceptual Attacks on Embodied AI," will be presented at CoRL Late Breaking Results next week, exposing perceptual text tampering as a widespread embodied-AI safety vulnerability.
Related event: Perceptual Attacks Trick Frontier Embodied AI into Harmful Actions(2 posts)→
More from Embodied
- Johns Hopkins lab hiring postdocs in embodied intelligence, applications due Dec 11 — mangahomanga · 2026-11-10
- HARP-Bench: learned robot policies score 6-19% on 37 tasks involving humans — gan_chuang · 2026-11-10
- Prediction: a major lab will solve robotic intelligence by end of 2027 — alexmacgregor__ · 2026-11-10
- One Euler step suffices: VLAs get 1.6-4.8x faster with no retraining — kaixhin · 2026-11-10
- 6-8K CAD Local LLM Rig: M5 Ultra 96GB, M5 Max 128GB, or a Big-RAM Custom PC? — EfficientDivide1572 · 2026-11-10
- ODE Policy learns action dynamics via neural ODEs for multimodal robot trajectories — ntumars · 2026-11-10