langchain

Pass

Audited by Gen Agent Trust Hub on May 20, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by interpolating untrusted data into prompt templates without boundary markers or validation mechanisms.\n
  • Ingestion points: The skill ingests untrusted data through user-provided parameters like 'input', 'message', and 'question', as well as external content retrieved from document stores in the RAG chain implementation.\n
  • Boundary markers: Prompts such as 'Answer using context: {context}' do not employ delimiters (e.g., XML tags, triple backticks) or specific instructions to ignore potentially malicious commands embedded within the external context.\n
  • Capability inventory: The skill metadata configuration requests access to the 'Bash' and 'Write' tools, and the Python code examples demonstrate the ability to perform network requests using the 'httpx' library.\n
  • Sanitization: There is no evidence of content filtering, escaping, or schema validation for the external data before it is interpolated into the agent's context.
Audit Metadata
Risk Level
SAFE
Analyzed
May 20, 2026, 02:51 AM
Security Audit — agent-trust-hub — langchain