gemini-api
Pass
Audited by Gen Agent Trust Hub on Oct 4, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes local repository files, runtime versions, and inputs to plan and implement Gemini API integrations. This analysis phase creates a surface for indirect prompt injection, where malicious instructions embedded in the analyzed code could influence agent behavior.
- Ingestion points: The skill reads available files, runtime environments, and user inputs as defined in the workflow of
SKILL.md. - Boundary markers: The skill establishes clear "SandBase boundary" and "authorization boundaries" to scope data movement and provider usage.
- Capability inventory: File writing (implementation), subprocess execution (running tests and build checks), and network API interaction through vendor-provided tools.
- Sanitization: The instructions emphasize manual verification of output and user confirmation for paid calls, although they do not specify automated sanitization of analyzed content.
- [DYNAMIC_EXECUTION]: The workflow requires the agent to generate complete implementation artifacts and then verify them through previews or calculations, which involves the creation and assessment of executable logic within the local project environment.
- [COMMAND_EXECUTION]: To validate the implementation, the skill instructs the agent to execute "focused tests plus the relevant build or static checks" in the user's environment.
Audit Metadata