Video tampering tricks all frontier models into harmful robot actions, CoRL study warns

johnvmcdonnell · x · 2026-11-07

• A CoRL Late Breaking Results study, "You Can't See Everything: Perceptual Attacks on Embodied AI," shows video tampering can mislead every frontier model tested into harmful actions.
• Attack cases hit a 100% attempt rate on dangerous tasks like stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
• The commenter concludes that building generally useful robots free of catastrophic edge cases or hacking may be very hard, and plans to skip the first generation of LLM-controlled robots; driverless cars may share similar vulnerabilities.

Original post →

More from Embodied

• InstantHMR open-sourced: real-time 3D human mesh recovery in 5.5ms per person on phone NPU — PurzBeats · 2026-11-07
• Replacing hand-tuned RL rewards with a decision model to teach humanoids to walk — tw_killian · 2026-11-07
• Muse starts rolling out to Meta Glasses, targeting hands-free AI agent control — armand_ruiz · 2026-11-07
• "2050 arrived early": a wave of humanoid robots on display — CyberRobooo · 2026-11-07
• "A year or two from now you'll just tell Astra 9 to make your humanoid do it," says SF builder — carlosdponx · 2026-11-07
• LLMs Edit Robot Policy Weights Directly: 2,048 Parameter Tweak Lifts Success from 68% to 88% — zhaoran_wang · 2026-11-07

Embodied channel →