pia-generation
Pass
Audited by Gen Agent Trust Hub on Sep 20, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill operates on an attack surface that allows for indirect prompt injection by aggregating untrusted data and performing file operations.
- Ingestion points: The skill ingests user-provided product requirements (PRDs), potentially including external links, and reads previous analysis outputs from the
~/.claude/plugins/config/directory. - Boundary markers: It employs mandatory source-tracking tags such as
[法条原文],[联网检索 — 需复核], and[用户提供]to distinguish between legislative text, search results, and user input. - Capability inventory: The agent writes assessment results to the local filesystem (
~/.claude/plugins/config/claude-for-legal-zh/privacy-legal/matters/) and invokes external tools for legal retrieval and web searching. - Sanitization: The instructions include a 'Prohibit Silent Supplementation' rule that prevents the agent from filling in gaps with training data without user verification, though it does not explicitly sanitize the content of ingested external documents.
Audit Metadata