Scott & Mark Learn To...Fool's Gold

Why watch

Scott Hanselman and Mark Russinovich discuss Mark's research on Fool's Gold, a novel AI defense technique that trains models to provide deliberately flawed outputs when safety guardrails are bypassed. The episode explores the security risks of open-weight AI models and whether this approach impacts legitimate model behavior. Ideal for developers, security researchers, and AI engineers interested in the evolving landscape of AI safety and model defenses.