maintain-verification-skill
Warn
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and processing untrusted data from the project's source code and verification skill files to detect documentation drift and generate verification recipes.
- Ingestion points: Project-local verification skill directory (SKILL.md, features/ directory, harness scripts) and product source code files accessed during the 'Source wave' phase.
- Boundary markers: The instructions do not specify any delimiters or safety warnings to distinguish between project data and agent instructions.
- Capability inventory: The skill can launch subagents, write to the skill's local directory, execute shell commands (via harness scripts), and create GitHub Pull Requests.
- Sanitization: There is no mention of sanitizing, escaping, or validating the ingested content before it is used to influence subagent tasks or update documentation.
- [DYNAMIC_EXECUTION]: The skill maintains and executes 'harness scripts' that are part of the project environment. It is explicitly authorized to modify these scripts and then execute them to verify application behavior.
- Evidence: Step 5 ('Triage') instructs the agent to fix 'harness gaps' by making scripts executable and updating documentation. Step 4 ('Live pass') then executes these scripts to exercise features, effectively running locally modified or generated code.
- [COMMAND_EXECUTION]: The skill performs active command execution during its 'Live pass' phase to drive application instances (servers, UIs, or CLIs).
- Evidence: Step 4 describes a launch model involving serial driving of long-lived instances or fresh sessions for CLIs, which inherently requires the execution of shell commands and utility scripts.
Audit Metadata