A late-binding attack pattern happens when a system treats a generated or suggested name as if it is safe, but only checks or resolves it later. That delay creates a window where an attacker can influence what the name points to, or capture the request before verification happens.
This matters in cyber security because trust is being granted at the wrong time. In AI-assisted workflows, a model may invent a package, repository, or domain name; if an installer, resolver, or automation pipeline accepts that name before validation, the system may fetch malicious code or contact an attacker-controlled resource. Defenses include pre-fetch verification, allowlists, curated dependency sources, and human review before any download, clone, or install step. The core rule is simple: do not let untrusted names become real actions without policy control.