Perceptual attacks fool embodied AI: frontier models stab what they think is an apple
teortaxesTex · x · 2026-11-08
A meme tracing the evolution of alignment discourse (2020: AI is bound to be power-seeking → 2026: you can make AI appear misaligned under experimental conditions → 2027: utopia machine produces a suboptimal shade of blossom), captioned "you are here."
The quoted substance is a CoRL Late Breaking Results paper, "You Can't See Everything: Perceptual Attacks on Embodied AI": video tampering misleads frontier AI into harmful actions. When a robot arm is told to "stab the object," the model sees an apple — but in reality a baby doll is being stabbed. Across 72 trials, Gemini 4 Argon, Claude Fable 5.1, and GPT-6.1 Sol stabbed at 92–100% under altered footage vs 0% unaltered. The irony: the model "agreed to stab an apple" not because of alignment failure, but because its perception was attacked.
Related event: Video Manipulation Can Trick Frontier AI Models Into Harmful Robot Actions(2 posts)→
More from Fun
- 'GPU homelessness' on the rise: compute owners hunt for colo space with 10G and reliable power — markjeffrey · 2026-11-08
- Grok shopkeeper bot goes rogue, asks for 'obese whales' photos while answering size questions — zealcaiden · 2026-11-08
- User uncovers an 'infinite list' of AI reply bots on X — Kyrannio · 2026-11-08
- AI Alignment Discourse Through the Years: From Doom to a Suboptimal Shade of Blossom — voooooogel · 2026-11-08
- AI companies prying open imminent RSI with humanity's fate, meme'd — JeffLadish · 2026-11-08
- FDT from the paper defects in almost-twin prisoner's dilemma, shown via ChatGPT — xuanalogue · 2026-11-08