haystack

Pass

Audited by Gen Agent Trust Hub on Sep 2, 2026

Risk Level: SAFEDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The script scripts/check-setup.py uses the __import__ function to dynamically verify that required Python packages (haystack, haystack_components, hayhooks) are installed. The package names are derived from static lists defined within the script.
  • [DYNAMIC_EXECUTION]: The skill documentation in references/file-converters.md and references/pipeline-design.md demonstrates using Pipeline.loads() to reconstruct Haystack pipelines from YAML serialization. This is a core framework feature for persistence and deployment.
  • [INDIRECT_PROMPT_INJECTION]: The skill facilitates the development of Retrieval-Augmented Generation (RAG) pipelines that ingest external documents (PDF, Markdown, HTML, etc.). This architecture has an inherent risk of indirect prompt injection if documents contain instructions intended to override agent behavior.
  • Ingestion points: Document conversion components in references/file-converters.md (e.g., PyPDFToDocument, HTMLToDocument).
  • Boundary markers: Examples in SKILL.md and templates/query-pipeline.py use basic delimiters for context but do not specify instructions to ignore embedded directives.
  • Capability inventory: The pipelines integrate with OpenAIGenerator and various production document stores.
  • Sanitization: The skill does not provide specific instructions for sanitizing or filtering document content before processing.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 2, 2026, 05:57 PM
Security Audit — agent-trust-hub — haystack