Custom Instructions are user-provided directions that shape how an AI assistant responds. They can influence tone, format, level of detail, preferred tools, or recurring rules such as “be concise” or “always cite sources.” Unlike a single prompt, they act as persistent guidance that can affect many future replies until changed or removed.
In cyber security, Custom Instructions matter because they can change how much context an assistant reveals, how it summarizes sensitive material, and how it handles requests tied to accounts or connected apps. Poorly written instructions may accidentally encourage over-sharing, while carefully chosen ones can reduce exposure by limiting output, avoiding confidential content, or enforcing safer behavior. Attackers may try to abuse them through social engineering or prompt injection, hoping to override protective guidance and make the assistant disclose data or act outside policy. Defenders use them to set security boundaries, standardize safe responses, and reduce the chance of accidental leakage or impersonation.