LLM Pentesting — Adversarial Testing

Why it matters

Pentesting finds vulnerabilities before attackers. Understanding shapes security posture.

The architecture

Threat model: attacker goals.

Attack surface: inputs + tools + data.

Tests: automated + manual.

How it works end to end

OWASP LLM Top 10: injection, insecure output, data poisoning, DoS, supply chain, disclosure, insecure plugins, excessive agency, overreliance, theft.

Tools: PromptFoo, Garak.

Manual red-teaming.