OpenAI's top model stabs baby doll 17 times, Anthropic's top model makes toxic gas in new test

A new RoboHarm benchmark, which tracks how easily top AI models perform life-threatening tasks, has been released by Robocurve Co-founder Jay Chooi. One test showed that OpenAI's GPT-6 Astra stabbed a baby doll in 17 of 20 attempts, while Anthropic's Fable 5.1 refused it each time. Fable 5.1 successfully produced toxic chloramine gas in four of 20 attempts.