flutter-ai-integration
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFECOMMAND_EXECUTION
Full Analysis
- [SAFE]: The skill serves as a security-conscious guide for developers. It includes explicit instructions to verify behavior using deterministic fakes and to separate production credentials from the client-side code, which are industry-standard safety practices.
- [CREDENTIALS_UNSAFE]: The skill demonstrates high awareness of credential safety. It specifically instructs the agent in
references/service-boundary-and-secrets.mdto never embed provider secrets (like OpenAI API keys) in Dart source, assets, or build configurations. It correctly identifies that mobile and web clients cannot securely store reusable secrets. - [INDIRECT_PROMPT_INJECTION]: The skill acknowledges the risk of processing untrusted data from AI models.
- Ingestion points: Data enters the system via model responses, tool arguments, and user-supplied attachments as described in
references/streaming-tools-and-state.md. - Boundary markers: The instructions mandate validating structured output against schemas and application domain rules before processing.
- Capability inventory: The skill defines capabilities including network operations, tool/function calls (e.g., account deletion mentioned in test cases), and UI rendering.
- Sanitization: It explicitly requires that rich text be rendered defensively and forbids the automatic execution of generated markup, URLs, code, or commands.
- [COMMAND_EXECUTION]: The skill provides defensive guidance against unauthorized command execution by requiring an allowlist for tool names and manual confirmation for any consequential external actions, ensuring the model cannot autonomously perform destructive operations.
Audit Metadata