ai-observability-promptfoo

Pass

Audited by Gen Agent Trust Hub on Apr 7, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill recommends using npx promptfoo@latest to execute the evaluation framework directly from the package registry.
  • [COMMAND_EXECUTION]: Support for custom javascript and python assertion types allows running arbitrary code on the host system to validate model outputs.
  • [DATA_EXFILTRATION]: Documents the --share flag which uploads evaluation results to the promptfoo.dev cloud service for sharing, which may include sensitive prompt or output data.
  • [PROMPT_INJECTION]: The skill identifies an indirect prompt injection attack surface where untrusted data from LLM outputs is processed by user-defined code. Ingestion points: model output in javascript or python assertions (e.g., examples/model-graded.md). Boundary markers: None. Capability inventory: Execution of local scripts via assertions. Sanitization: No built-in sanitization for output data before processing by custom logic.
  • [EXTERNAL_DOWNLOADS]: Fetches the promptfoo package from the npm registry during execution and uses GitHub Actions for CI/CD integration.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 7, 2026, 01:31 AM