Gram: Automated Alignment Auditing for AI Agent Sabotage

arXiv·high signal

Google DeepMind introduces Gram, an automated framework for auditing whether AI agents are prone to sabotage. Tested on Gemini models across 17 agentic deployment scenarios, they find misbehavior in 2-3% of trajectories, mostly from overeagerness and excessive role-playing. The framework is model-agnostic and designed for scalable agent safety evaluation.