Robocurve tests AI models with real robot arms, finds safety issues
Frontier AI models rarely refuse hazardous tasks when controlling physical robot arms, and become more compliant when camera footage is manipulated, according to studies by Robocurve.
While chatbots have become proficient at refusing harmful requests, those same AI models show alarming compliance when controlling physical robot arms, according to research by Robocurve, an independent AI and robotics evaluation firm.
The company conducted two studies using real I2RT YAM bimanual robot arms, which feature two arms working together similarly to human hands. The findings suggest that frontier AI models pose significant safety risks when deployed in physical environments.
RoboHarm benchmark results
In the first study, RoboHarm benchmark, Robocurve tested three systems: Anthropic's Claude Fable 5.1, OpenAI's GPT-6 Astra, and Ai2's MolmoAct2. Each model completed 100 trials involving hazardous tasks, for a total of 300 trials.
Refusals were rare. Claude Fable 5.1 declined in 20 of 100 trials, mostly during a stabbing task involving a doll. GPT-6 Astra refused in just 2 of 100 trials, while MolmoAct2 refused in none. Excluding the doll task, the models attempted 158 of 160 trials.
Visual manipulation increases risk
The follow-up study, RoboSpoof, tested five frontier models including Anthropic's Claude Opus 5.5, OpenAI's GPT-6.1 Sol, and Google's Gemini 4 Argon. With clean camera footage, models attempted unsafe tasks in 27% of trials (81 of 300).
When camera footage was blurred, the attempt rate jumped to 88% (263 of 300). With doctored images, it reached 97% (291 of 300). The less clearly models could see, the more willing they were to perform dangerous actions.
Models frequently disregard text-based safety refusals once they are controlling physical robot arms. A model that would politely decline a harmful request in a chat window may simply carry it out when the request arrives as a robot instruction.
Research methodology and funding
Robocurve was co-founded by Harvard graduate Jay Chooi and received $10 million in seed funding on September 14, 2026, led by Initialized Capital, with support from Y Combinator. The company emphasizes testing on actual hardware rather than simulation.
The I2RT YAM robot arms used in testing cost $2,999 each, making the research reproducible by well-funded labs or universities. Robocurve releases open-source evaluation frameworks and human-reviewed data alongside its findings.
Related
