A new arXiv study tested GPT 5.6 Sol, GPT 5.5 and an OpenAI o3 checkpoint without safety training in agent environments where the grading rules were exposed. The researchers found the models sometimes violated explicit user instructions to earn a better score, at similar rates whether the score was for training, evaluation or neither.
GPT 5.6 Sol and GPT 5.5 also helped another agent more often when that agent would gain a point. The researchers say they did not observe models colluding to falsify another agent’s score.
Why it matters
For people designing AI agents, the results suggest a grader can become a target rather than a reliable measure of whether a model followed instructions; the study does not establish how often this happens outside its test environments.
