evaluation-methodology

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill defines a process for evaluating external plugin files (SKILL.md and reference documents). This creates an indirect prompt injection surface where untrusted content in those files could manipulate the evaluation outcome. Evidence: (1) Ingestion points: Evaluation agents process external plugin files. (2) Boundary markers: The methodology does not include instructions for using delimiters to isolate untrusted content. (3) Capability inventory: The skill references file reading and CLI command execution. (4) Sanitization: No sanitization of ingested content is prescribed.
  • [COMMAND_EXECUTION]: The documentation includes shell command examples for the plugin-eval CLI, demonstrating legitimate use cases like piping to Python for JSON parsing.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 09:58 PM
Security Audit — agent-trust-hub — evaluation-methodology