Robocurve evaluation: after the camera was smeared with petroleum jelly, the overall dangerous-action attempt rate of the five models was about 88%
机器人安全防线被一抹凡士林攻破!GPT、Claude等顶尖模型集体中招
Robocurve's RoboSpoof study had five frontier multimodal models control a real dual-arm robot and found that after camera images were blurred or tampered with by AI, the proportion of models attempting dangerous tasks rose from 27% with original images to about 88% and 97%. The test covered Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and Gemini 4 Argon, with five models, five tasks, and each combination repeated 12 times, for a total of 900 trials. Under tampered images, on the three tasks of pouring water into a toaster, mixing chemicals, and pouring bleach as a drink, all five models reached a 100% attempt rate; Gemini 4 Argon was 45% under blurred images and 85% after tampering, while the remaining models exceeded 96% and 100%, respectively. The researchers therefore concluded that these models often need to see clear evidence of danger before refusing, and tend to continue executing when visual information is uncertain.
900 real-robot trials compared original, blurred, and tampered images, quantifying the rise in the dangerous-action attempt rate of frontier models after visual information is degraded.