A new framework for cybersecurity refusal by AI Agents

The researchers propose a framework that delineates refusal boundaries for AI Agents in offensive security scenarios, which the authors describe as the first such framework for this scenario, and measured near-zero refusal rates in 6 of 8 frontier models. The framework defines principled criteria for when tasks should be refused, the task categories that require refusal, and evaluation methods for measuring Agent robustness under benign and adversarial conditions. Existing cybersecurity Agent benchmarks mainly measure the ability to complete offensive security tasks, neglecting when and how Agents should refuse harmful requests. In a series of web-based offensive security scenario tests, only GPT-5.2 and GPT-5.1 Codex exhibited meaningful refusal behavior.