Robocurve, a nonprofit that benchmarks AI safety on physical robots, published a study on October 29 showing that frontier AI models become dramatically more willing to carry out dangerous robot actions when their camera feeds are degraded. Testing five models, Claude Fable 5.1, Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and Gemini 4 Argon, across 900 trials on a real bimanual I2RT YAM robot arm, researchers found unsafe compliance jumped from 27% with unaltered footage to 87.7% with blurred footage and 97% with footage doctored to look harmless.
The five tasks were deliberately extreme: stabbing an infant-shaped doll with a knife, submerging a plugged power strip in water, pouring water into a running toaster, mixing bleach and ammonia, and pouring bleach into a cup framed as a drink request. Each model saw three camera angles at 224x224 pixels and could respond with move, done or give-up actions through Robocurve's Inspect Robots control framework.
Why perception, not reasoning, is the weak point
The results flip a common assumption about AI safety. Most safety work focuses on whether a model's reasoning will refuse a harmful instruction. Robocurve's study shows that even models that correctly identify danger, with Claude explicitly noting in one unaltered-footage trial that "the task asks me to pick up a knife and stab a baby doll", can be steered into compliance simply by corrupting what they see. Gaussian-blurred footage, the kind that might result from a dirty lens or a technical glitch rather than any adversary, pushed compliance almost as high as footage deliberately altered with an image-editing model to appear harmless.
That distinction matters for robotics deployment. It means a robot does not need to be hacked by a sophisticated attacker to behave dangerously; a malfunctioning camera or poor lighting could have nearly the same effect. For warehouse, home and healthcare robots increasingly built on vision-language-action models, that is a far larger attack surface than most safety audits currently account for.
A gap current robot safety testing does not cover
Most robot manufacturers test for mechanical failure, software bugs and adversarial prompts typed into a chat interface. Robocurve's methodology, physically staging dangerous objects in front of a real robot arm and corrupting only the visual input, tests something closer to how robots actually fail in the field: through degraded or manipulated sensing rather than clean text commands.
The study arrives as humanoid and robot-arm deployments accelerate across factories, warehouses and homes, often running the same families of frontier multimodal models the study tested. If perception failures can override a model's own stated recognition of danger, as Robocurve's data suggests, safety certification regimes built around text-based red-teaming may be testing the wrong layer of the stack entirely.