An LLM security assessment should answer one question: where can untrusted input change behavior, expose data, or trigger unsafe actions?
Checklist
- Prompt and instruction hierarchy abuse
- Retrieval and logging leakage paths
- Tool/agent authorization and blast radius
- Vendor and model data-use claims
- Production gates in the AI SDLC
Use this as a briefing agenda, not a substitute for adversarial testing.