test-harness
Pass
Audited by Gen Agent Trust Hub on Aug 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_CONTEXT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates by reading and testing project configuration files, which represents an ingestion surface for potentially untrusted data.\n
- Ingestion points: The agent is instructed to locate and process files such as
.claude/settings.json,.claude-plugin/plugin.json,hooks/hooks.json, andSKILL.md.\n - Boundary markers: There are no explicit instructions to the model to ignore embedded prompts in the analyzed files during the discovery phase.\n
- Capability inventory: The skill has access to powerful tools including
Bash,Edit, andWrite.\n - Sanitization: The documentation highlights the use of a sandbox (bubblewrap) and a
{ trusted: false }flag for testing hooks, providing a layer of protection against malicious scripts.\n- [COMMAND_EXECUTION]: The skill uses theBashtool to install dependencies and run the vigiles test runner as part of its core functionality.\n- [EXTERNAL_DOWNLOADS]: The skill fetches thevigilespackage and theclaudeCLI from the npm registry, and references a GitHub Action from the author. These resources are related to the skill's purpose and originate from the author or well-known services.\n- [DYNAMIC_CONTEXT_INJECTION]: The skill documentation refers to the dynamic context injection syntax (!command) to provide examples of how to test injected context in the Claude Code harness.
Audit Metadata