skill-eval
Pass
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill executes external command-line tools (such as
gitand various AI agent adapters likeclaudeorcursor-agent) via Node.jschild_processmodules to facilitate its evaluation workflow. This is a core part of its intended functionality as a developer tool. - Evidence includes the use of
spawnandexecFileSyncinscripts/run-harness-eval.mjsandscripts/compute-eval-verdict.mjsto execute adapters and verify repository state. - [PROMPT_INJECTION]: As a tool that processes the output of other AI agents, it is naturally exposed to indirect prompt injection from evaluated artifacts.
- Evidence: The skill reads and evaluates raw event streams and artifact files produced by sub-processes.
- Sanitization: The skill mitigates this risk by using strict evidence verification with SHA-256 hashes, enforcing path isolation within the
.skill-eval-runs/directory, and instructing the agent to evaluate results based on deterministic assertions rather than surface-level formatting or claimed procedures.
Audit Metadata