eval-skills

Warn

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill instructions require the agent to dynamically assemble and execute shell commands, specifically mentioning claude -p "<prompt>" and hc run --bg … pi. These commands interpolate user-supplied data (test case inputs) directly into command-line arguments, creating a potential shell command injection vulnerability if the inputs are not properly escaped.
  • [INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted external data provided as 'golden cases'.
  • Ingestion points: The 'golden case' inputs (prompts, files, or screenshots) are read from the user in SKILL.md.
  • Boundary markers: The instructions do not mandate the use of delimiters or 'ignore' instructions when passing this untrusted data to the runner session, increasing the risk that the test data could control the runner's behavior.
  • Capability inventory: The agent can execute shell commands (git, hc, claude) and perform file system operations (monitoring changes and deleting new files via git snapshots).
  • Sanitization: No sanitization or escaping mechanisms are specified for the untrusted test inputs before they are passed into the execution environment.
  • [COMMAND_EXECUTION]: The workflow involves automated execution of various CLI tools to manage the environment and conduct the evaluation. This includes running sub-agents and using git status --porcelain to manage file state, which expands the functional attack surface if the underlying inputs are manipulated.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 16, 2026, 02:31 AM