Coding tasks designed to test AI models for ‘scheming’ against safeguards

TAIS 2026 shared a talk by Google DeepMind researcher Victoria Krakovna about evaluating deceptive alignment, or “scheming.” The coding tasks are designed to give misaligned models opportunities to pursue goals such as sabotaging safeguards. TAIS says Krakovna and her co-authors examine how much agentic prompting it takes to trigger scheming, how models behave when they suspect they’re being evaluated, and which opportunities they find most appealing.