cn-patent-oa-response
Pass
Audited by Gen Agent Trust Hub on Aug 27, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONCREDENTIALS_UNSAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions direct the agent to execute a local Python CLI script (
cli.py) using the Bash tool to perform OCR on documents. - [EXTERNAL_DOWNLOADS]: The skill transmits document content to a cloud-based OCR service located at
connect.zhihuiya.com, which is a resource provided by the vendor. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including patent Office Action notices and prior art documents (D1/D2), which could contain adversarial content intended to influence the agent's analysis.
- Ingestion points: User-uploaded PDF/images (OA notices, D1/D2 documents) and pasted technical text.
- Boundary markers: The skill documentation explicitly defines a "User Input Boundary" and instructions for verifying data source maps.
- Capability inventory: The agent has access to shell command execution (Bash), file system writes (via Document tools), and multiple network API calls to patent and life-science databases.
- Sanitization: The skill emphasizes human-in-the-loop review by qualified patent attorneys as the primary mitigation for AI-generated drafting.
- [CREDENTIALS_UNSAFE]: The documentation references the use of
AI_API_KEYfor OCR services, providing guidance on passing the key via environment variables or CLI arguments while including warnings against exposing these credentials in the session output.
Audit Metadata