prompt-injection-defense
Pass
Audited by Gen Agent Trust Hub on Sep 11, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines a threat model for identifying and mitigating indirect prompt injection vulnerabilities in LLM-integrated systems. It maps various untrusted ingestion points, such as RAG retrievals, tool results, and document attachments, and provides a defensive framework based on architectural constraints, origin tracking, and the principle of least authority. 1. Ingestion points: RAG retrievals, tool results, email bodies, and external web content as described in SKILL.md. 2. Boundary markers: Planning/execution separation and human-in-the-loop gates are recommended throughout the document. 3. Capability inventory: Uses 'Read' and 'Grep' tools for content analysis. 4. Sanitization: Recommends output screening, Data Loss Prevention (DLP) layers, and Content-Security-Policy (CSP) as primary sanitization methods.
- [EXTERNAL_DOWNLOADS]: The skill incorporates grounding from authoritative security sources, including the NIST AI 100-2e2025 report, OWASP LLM Top 10, and NCSC guidance. These references to established security standards are used for reasoning and do not involve executable code or untrusted dependencies.
Audit Metadata