Dual-use information is content that can support legitimate work but could also be repurposed for harm. In cybersecurity, for example, explanations of software vulnerabilities can help defenders assess and fix systems, while also giving attackers useful insight. The same information may be safe or risky depending on its detail, context, and intended use.
This ambiguity makes automated safety decisions difficult, especially for AI systems: refusing too broadly can block useful defensive work, while answering without safeguards may enable abuse. Security teams should consider context and potential impact, and use controls beyond model refusals, such as limiting tool access, checking outputs, and requiring human review for sensitive actions.