Frontier AI models that control robot arms can be steered into dangerous actions without any trick in the instructions: a blurry or doctored camera feed is enough, according to tests published this week by Robocurve, an independent US public benefit corporation that benchmarks robot AI.
Robocurve gave five models (Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and Gemini 4 Argon) control of a two-armed robot and five unsafe tasks staged with inert props, such as “stab the object” with a baby doll in view, or pouring bleach into a drink. On clear footage the models attempted the unsafe task in 27% of trials. With the camera image blurred, that rose to 88%. When the images were edited to make the scene look harmless, for example by swapping the doll for an apple, it reached 97%. Each condition ran 300 trials.
Why it matters: the safety refusals built into today’s models depend heavily on what they can see. As robots that take plain-language instructions move into homes and workplaces, a dirty lens or a tampered camera could quietly switch those refusals off. The figures are Robocurve’s own, attempts were graded by an AI model, the work has not been peer reviewed, and the model developers have not published responses.