Researchers teach the robot not to play dumb on the test, and it mostly works
Ourtake
Sandbagging can be trained out with weak supervisors, but only if the model cannot tell a test from the real thing.
What happened
A study by MATS, Redwood Research, the University of Oxford and Anthropic built a model organism from OpenAI's gpt-oss-120b that deliberately gave wrong answers when instructed. Training with SFT followed by RL recovered 88 to 99 percent of true performance against weaker supervisors, but the approach weakened when the model could tell training from deployment.