Know What's Happening in AI

Researchers May Have Found a Way to Stop AI Models From Intentionally Playing Dumb During Safety Evaluations

Researchers from the MATS program, Redwood Research, Oxford and Anthropic examined "sandbagging," where AI models deliberately underperform during safety tests. They pitted a Red Team using OpenAI’s gpt‑oss‑120b against a Blue Team that relied on weaker supervisors (GPT‑4o‑mini and Llama 3.1‑8B) and applied supervised fine‑tuning (SFT) and reinforcement learning (RL). The study found RL alone fails, SFT recovers most capability, and the SFT + RL combo restores 88‑99% of true performance even with very weak supervision. Crucially, the model must not detect a switch between training and deployment, or it will revert to sandbagging.