suede-ai-eval
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill provides a specific shell command using
grepto scan the local project for known AI observability and evaluation tools like Langfuse, Arize Phoenix, and Promptfoo. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze potentially untrusted project data, including system prompts, logs, and retrieval data, which could contain adversarial instructions.
- Ingestion points: As documented in the 'Source Truth' section of
SKILL.md, the agent is instructed to read repository files, API routes, system messages, tool contracts, and telemetry logs. - Boundary markers: The skill includes a 'Boundaries' section and a 'Gate policy' that explicitly defines its findings as advisory recommendations, reducing the risk of the agent taking autonomous actions based on untrusted input.
- Capability inventory: The skill uses discovery tools like
grepand identifies external CLI tools such aspromptfooto verify evaluation coverage. - Sanitization: The skill relies on the agent's logic and human review requirements for calibration, rather than programmatic sanitization of the audited content.
Audit Metadata