voice
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill creates and populates markdown files in a
voice/directory which are explicitly designed to be loaded and used as instructions by other content-generating skills (e.g.,/write,/sales). - Ingestion points: The skill reads from and writes to the
voice/directory in the local file system (SKILL.md, Steps 1 and 2). - Boundary markers: The skill does not define or implement delimiters (like XML tags or specific markers) to wrap the user-provided style content to prevent it from being interpreted as instructions by consuming models.
- Capability inventory: The skill is limited to directory listing and file creation/writing; it does not possess network access or high-privilege execution capabilities.
- Sanitization: The instructions explicitly state 'Don't paraphrase — use their exact words where possible,' which means user-provided content is stored without filtering or sanitization.
- [COMMAND_EXECUTION]: The skill includes a shell command to verify the project state.
- Evidence: A bash one-liner
ls voice/ 2>/dev/nullis used in Step 1 to check for the existence of the configuration directory. This is a standard, low-risk operation.
Audit Metadata