A model router is a platform layer that sends each AI request to one of several models based on cost, latency, accuracy, or policy. Instead of hard-coding one provider, teams can route simple tasks to cheaper models, sensitive tasks to approved models, and high-value tasks to stronger ones. This matters in cybersecurity because the router becomes a control point for where data goes, which model sees it, and what outputs are allowed to return.
In real attacks, a compromised or misconfigured router can leak prompts, route regulated data to an unauthorized model, or let attackers steer traffic toward weaker defenses. Prompt injection can also try to manipulate routing rules or escape policy checks. Defenders use allowlists, per-task policies, audit logs, secret redaction, and fallbacks that fail closed when a request is ambiguous. In secure AI stacks, the router is not just a cost tool; it is part of the trust boundary.