gs-create-verification-skill
Warn
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill generates a new executable agent skill (SKILL.md) containing shell commands and then immediately executes those commands to verify the generated skill's functionality. This 'write-then-run' pattern is a significant vector for executing unintended code.
- Evidence: Step 4 ('Prove the generated skill before handing it over') explicitly directs the agent to 'Run its own instructions end to end once: launch, doctor, drive ONE mapped feature...'.
- [INDIRECT_PROMPT_INJECTION]: The skill's logic depends on 'interviewing' the local codebase to extract 'the exact command that starts the app'. It treats the contents of the repository as trusted instructions for how to interact with the environment.
- Ingestion points: Local repository files including README.md, Makefile, package.json, and other project documentation as specified in Step 1 ('Interview the repo, not the user').
- Boundary markers: There are no instructions for the agent to sanitize or ignore potentially malicious embedded instructions in these files.
- Capability inventory: The agent possesses capabilities for file system writes (to create the skill directory and files) and arbitrary command execution (to run the 'launch' and 'drive' steps).
- Sanitization: The instructions lack a validation step to ensure extracted commands are safe before they are written to the new skill and executed.
- [COMMAND_EXECUTION]: The skill facilitates the execution of arbitrary development commands, which may include installation of third-party dependencies or compilation of binaries from the source code.
- Evidence: Step 2 ('Launch' section) instructs the agent to determine the start command, including potential steps to 'build the binary (or install deps)'.
Audit Metadata