retrieval-strategy
Pass
Audited by Gen Agent Trust Hub on Oct 2, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill provides a 'Contextual compression' pattern in Section 6 of SKILL.md which interpolates retrieved document chunks directly into an LLM prompt, creating a surface for indirect instructions to influence the agent.\n
- Ingestion points: Untrusted document content is ingested via the
chunk_textvariable in thecompressfunction.\n - Boundary markers: The prompt template relies on simple text labels ('Passage:') rather than robust delimiters or explicit directives to ignore embedded instructions within the retrieved data.\n
- Capability inventory: The skill demonstrates the use of an LLM client (
client.messages.create) to process this untrusted data.\n - Sanitization: No sanitization or filtering logic is present to identify or escape instructions embedded within the retrieved text.
Audit Metadata