llm-pipeline
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to orchestrate pipelines that process untrusted external data, such as support tickets or user queries, creating an attack surface for indirect prompt injection.\n
- Ingestion points: External data processed by the pipeline, such as support tickets mentioned in
evals/cases.yaml.\n - Boundary markers: The instructions advocate for using Pydantic and JSON Schema for typed contracts between steps, which acts as a data boundary, though specific delimiters for the LLM prompts themselves are not mandated.\n
- Capability inventory: The skill manages LLM call chains and uses a router for automated failover between models.\n
- Sanitization: Recommends rigorous validation of structured outputs at each seam in the pipeline to prevent data poisoning between subsequent calls.\n- [COMMAND_EXECUTION]: The skill includes a
scripts/verify.shfile that executes local commands for static analysis of the project's code and configuration.\n - Evidence: The script uses standard utilities such as
find,grep,sed, andpython3to perform its checks.\n - This behavior is localized for development verification and does not involve remote code execution or processing untrusted network data.
Audit Metadata