VirtualCrime: Evaluating the Criminal Potential and Agentic Behaviors of Large Language Models via Sandbox Simulation

Large language models (LLMs) have shown strong capabilities in multi-step decision-making, and are increasingly integrated into real-world agentic applications. While existing safety evaluations focus on static interactions or refusal rates, it remains largely underexplored what level of threat these models pose once an adversary successfully assigns them a malicious role. To evaluate this agentic misuse risk, we propose VirtualCrime, a dynamic multi-agent sandbox framework that elicits and measures multi-step criminal behavior from LLMs. We design 40 diverse crime tasks within this framework, covering 11 realistic maps and 14 objectives such as theft, robbery, and kidnapping. We evaluate 8 strong LLMs and find that: (1) even with robust safety alignment, all evaluated models exhibit alarming agentic capabilities to generate and execute detailed multi-step criminal plans when placed in an adversarial red-teaming setting; (2) models frequently suffer from behavioral alignment failures, exhibiting a high propensity for over-harm by escalating to violent means far beyond what the task requires; (3) their most sophisticated criminal behaviors concentrate in social deception and team coordination rather than technical exploitation. Our work highlights a critical gap in current safety paradigms, emphasizing the urgent need for behavioral alignment in autonomous agent deployments.