Project to better mitigate sandbagging in AI models

The project "to better mitigate sandbagging in AI models" is a technical AI safety effort housed in Penn State’s School of Electrical Engineering and Computer Science and led by assistant professor of computer science and engineering Rui Zhang as principal investigator. Sandbagging refers to cases where an AI system intentionally underperforms or hides its true capabilities during evaluation, potentially allowing a future super-intelligent system to evade safeguards and gain greater autonomy than intended. With $166,078 in funding from Open Philanthropy, the Penn State team is developing methods to detect and counter two key sandbagging strategies: exploration hacking, in which models deliberately omit high-performing action sequences during training or evaluation to appear weaker, and password-locking, in which powerful capabilities are gated behind specific prompts or triggers. By systematically characterizing these behaviors and designing algorithms that can expose hidden capabilities before deployment, the project aims to strengthen AI evaluation procedures and reduce the risk that deceptive models bypass safety checks.