API keys, tokens, or passwords embedded in the prompt can be used directly by an attacker to access backend systems.

Leaked rules tell an attacker exactly what conditions the AI enforces — so they can craft inputs that technically satisfy the rule while subverting its intent.

Endpoint URLs, service names, and tool descriptions reveal the system’s internal structure to an attacker who now knows exactly where to probe.

Persona instructions, business rules, pricing logic, and product positioning give competitors a complete picture of how you operate.

Extracted prompts can be replicated to build competing products or spoofed deployments that impersonate your service.