ai-eval-ci
Pass
Audited by Gen Agent Trust Hub on Mar 13, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the Promptfoo evaluation framework using
npx promptfoo@latestand utilizes official GitHub Actions from the verifiedactionsorganization.\n- [PROMPT_INJECTION]: The custom evaluation logic in the TypeScript example (Strategy 2) presents a surface for indirect prompt injection by interpolating AI agent outputs into a judge prompt.\n - Ingestion points:
output(AI Output) andrubricvariables in thejudgefunction withinSKILL.md.\n - Boundary markers: Absent; untrusted AI output is concatenated directly into the prompt string.\n
- Capability inventory: The script executes LLM API calls via the
openailibrary to generate qualitative scores.\n - Sanitization: No sanitization or delimiter-based isolation is applied to the AI output before processing. This is documented as the intended functional pattern for LLM-as-judge evaluations.\n- [SAFE]: Sensitive API keys for OpenAI and Anthropic are accessed via GitHub Action secrets (
${{ secrets.OPENAI_API_KEY }}) rather than being hardcoded in scripts.
Audit Metadata