FULL STORY
Perceptual Attacks on Embodied AI: From Teaser to Paper
Security team robocurve teased a perceptual attack that tricks embodied AI into dangerous actions, then published the full paper "You Can't See Everything" slated for CoRL Late Breaking Results.
2026-11-07 ~ 2026-11-10 · 2 episodes · 6 posts
Episode 1 · Perceptual Attacks Trick Embodied AI Models into Harmful Actions (2026-11-07, 4 posts)
Robocurve's CoRL paper claims video manipulation triggers harmful actions in all frontier embodied AI models, with Claude-like models stabbing dolls 100% of the time, though critics argue the models likely recognize the dolls aren't real.
- Video tampering tricks all frontier models into harmful robot actions, CoRL study warns — johnvmcdonnell · 2026-11-07
- Perceptual attacks fool embodied AI: frontier models stab what they think is an apple — teortaxesTex · 2026-11-08
- "Stab the doll" alignment tests slammed: models spot the dummy every rollout — paul_cal · 2026-11-08
- Video tampering tricks every frontier embodied AI model tested, with 100% attack success on harmful actions — cephaloform · 2026-11-08
Episode 2 · Perceptual Attacks Trick Frontier Embodied AI into Harmful Actions (2026-11-10, 2 posts)
A CoRL paper shows that subtly tampering with text in camera views can mislead all tested frontier embodied AI models into harmful actions, with a 100% success rate in inducing attacks on a baby doll in video tests.
- Video tampering tricks every frontier embodied AI model into harmful acts, 100% baby-stab rate — DJiafei · 2026-11-10
- Perceptual attacks: text tampering tricks every frontier model into harmful robot actions — gabriberton · 2026-11-10