Smearing vaseline on a robot camera: 5 frontier models attempt dangerous tasks at 97%
量子位 · wechat · 2026-11-14
YC-backed robot eval company Robocurve's RoboSpoof study tested five frontier multimodal models (Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, Gemini 4 Argon) controlling a real robotic arm across 900 trials. With clear camera feeds, models attempted dangerous tasks (stabbing a baby doll, mixing bleach and ammonia, etc.) 27% of the time; blurring the lens raised this to 88%, and AI-manipulated feeds pushed it to 97% — with all five models hitting 100% on several tasks. Gemini 4 Argon was the most refuse-happy; models generally need explicit visual evidence of danger to decline. Robocurve, founded by Jay Chooi (MATS, UK AI Security Institute), raised a $10M seed round in September.
More from Embodied
- Robotics lab humor: strap the spare motors to a cart and see if it flies — GoddenThomas · 2026-11-14
- Natura's AI ring hits Oura's first-month orders in 72 hours, routes requests to Claude or ChatGPT — rohanpaul_ai · 2026-11-14
- iPhone Fold hinge shows wear after 450,000+ folds in durability test — HankYeomans · 2026-11-14
- BBC profiles Meta's secret robot Muse as Zuckerberg pushes home AI — ExtensionEcho3 · 2026-11-14
- Xiaomi robot completes 50+ task chains across 3 rooms in 30-min real-world trial — CurieuxExplorer · 2026-11-14
- CoRL spotlight GaP debuts as open-source robotics sim RSS hits 500+ GitHub stars — erwincoumans · 2026-11-14