giggle-voice-clone
Pass
Audited by Gen Agent Trust Hub on Sep 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill takes user-controlled inputs (reference audio URL, text for synthesis, and a voice identifier) and interpolates them into a bash command executed via the
exectool, creating a surface for command injection or instructions that could manipulate the shell process.\n - Ingestion points: The
--audio-url,--text, and--voice-idparameters in the command example withinSKILL.mdare directly populated from user input.\n - Boundary markers: No specific boundary markers or 'ignore embedded instructions' warnings are provided to the agent to prevent it from executing malicious payloads contained within the synthesized text or URL parameters.\n
- Capability inventory: The skill invokes the
exectool to run a local Python script (scripts/voice_clone_api.py) and uses therequestslibrary for network access to the vendor's API.\n - Sanitization: While the Python script itself uses
argparsefor internal parsing, there is no validation or shell-escaping of the command string performed at the prompt interpolation layer where the command is constructed by the agent.
Audit Metadata