AI Security Testing 101: Prompt Injection and Indirect Prompt Injection

AI security testing focuses on vulnerabilities like prompt injection, as outlined in the OWASP AI Testing Guide's AITG-APP category. Direct Prompt Injection occurs when attacker-controlled input alters AI application behavior. Beyond input sanitization, prompt injection enables AI agent impersonation, where agents use valid credentials for unauthorized purposes. To mitigate these identity risks, security controls such as delegated tokens with act claims, per-call audience binding, sender-constrained tokens, and external policy checks are essential for protecting agentic AI.