maintain-verification-skill
Pass
Audited by Gen Agent Trust Hub on Oct 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: Phase 4 (Live pass) of the skill instructions directs the agent to execute 'recipe blocks' extracted from Markdown files within the project's skill directory. This involves executing arbitrary shell commands or scripts that are documented as verification steps.
- [DYNAMIC_EXECUTION]: The skill performs dynamic execution of local content by extracting Markdown code blocks at runtime and executing them in a documented order (Phase 4). This pattern is inherent to the skill's purpose but presents a risk if the source files are untrusted.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from the project's source code and existing feature maps, which may be controlled by third parties in collaborative environments.
- Ingestion points: Reads project source code (Phase 2) and project-local skill directories like
.claude/skills/or.agents/skills/(Phase 0). - Boundary markers: The instructions do not specify any boundary markers or instructions to ignore embedded prompts when the agent (or its subagents) reads source code or feature files.
- Capability inventory: The skill possesses capabilities for file system read/write, local command execution (Phase 4), and version control operations such as creating Pull Requests (Phase 6).
- Sanitization: There is no evidence of sanitization, validation, or escaping of the content read from source code or documentation before it is used to influence the agent's verification recipes.
Audit Metadata