LLM Agent Security Design Guide: From Prompt Injection to Secure Deployment
How to apply least privilege to LLM agents: minimize tools, drop open-ended tools, narrow service-account permissions, run in the user's context, and require approval for risky actions. Includes an OWASP LLM06 (Excessive Agency) checklist, server-side tool authorization code, and prompt-injection and guardrail defenses.
Jul 23
LLM Agent Security Architecture: A Design Guide for Runtime Guardrails Against Prompt Injection and Jailbreak Attacks
LLM agent vulnerabilities arise at the system integration layer, not from model performance. This guide presents a methodology for designing multi-layered runtime security guardrails—including input filtering, intent verification, and outpu
Jun 8