ai-model-wechat
Pass
Audited by Gen Agent Trust Hub on Nov 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data, including user-provided prompts and image URLs, through vision-capable AI models. This creates a surface for indirect prompt injection where malicious content in an image or prompt could influence agent behavior.
- Ingestion points: Data enters the context via the
messagesarray ingenerateTextandstreamTextcalls, and throughimageUrlparameters (found inSKILL.md). - Boundary markers: The instructions recommend using specific prompt constraints (e.g., "Output JSON only. No explanation") to delimit model output, though these are not cryptographic boundaries.
- Capability inventory: The skill utilizes powerful capabilities including database writes (
db.collection().add), file uploads (wx.cloud.uploadFile), and environment configuration management viacallCloudApi(specificallyUpdateAIModelandCreateAIModelactions). - Sanitization: The skill provides a defensive parsing pattern for model responses using regex and
JSON.parse. Critically, it advocates for a human-in-the-loop pattern, instructing developers to render extracted data for user review and correction before committing it to a database.
Audit Metadata