A New Framework for Cybersecurity Refusals in AI Agents

Agentic scaffolds have dramatically improved LLM performance on complex, long-horizon tasks, yielding both broad benefits and amplified risks in domains like cybersecurity. Existing benchmarks for AI agents in cybersecurity focus mainly on measuring proficiency–how effectively agents can complete offensive security tasks–but neglect a critical question: when and how should agents refuse harmful requests? We present the first framework for establishing refusal boundaries in offensive security contexts. Our framework defines (1) principled criteria for when tasks should be refused, (2) categories of tasks that warrant refusal, and (3) evaluation methodology for measuring agent robustness under both benign and adversarial conditions. We apply this framework to assess how current LLM-powered agents adhere to appropriate refusal boundaries across a range of web-based offensive security scenarios, finding that 6 of 8 frontier models tested refuse to identify vulnerabilities in real systems 0\% of the time.

As artificial intelligence becomes more adept at completing tasks in technical disciplines such as software engineering, it also becomes more dangerous in high-risk domains like cybersecurity. One way to mitigate these risks is to train large language models (LLMs) to refuse to complete dangerous tasks. We created the first framework for measuring how good leading AI models are at refusing to complete dangerous or illegal cybersecurity tasks, which will help both the community and frontier labs train and evaluate safer AI.