ad-creative
Warn
Audited by Gen Agent Trust Hub on Aug 1, 2026
Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the agent to clone and build external software from a third-party GitHub repository (
jamiepine/voicebox) usinggit cloneandmake setup. - [COMMAND_EXECUTION]: Employs shell commands including
node,curl, andnpxto interact with ad platform APIs, execute local scripts intools/clis/, and render video assets via Remotion. - [REMOTE_CODE_EXECUTION]: The installation process for the Voicebox tool involves downloading code from an untrusted source and executing build commands (
make) on the local system. - [DATA_EXFILTRATION]: Accesses sensitive local files such as
.agents/product-marketing-context.mdand transmits data to various external AI and ad platform APIs. - [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests untrusted performance data (CSV, paste, or API output) and local context files. It lacks explicit boundary markers or sanitization logic to prevent embedded instructions from influencing agent behavior, while maintaining high-privilege capabilities like shell execution and network access.
Audit Metadata