diligence-issue-extraction

Pass

Audited by Gen Agent Trust Hub on May 15, 2026

Risk Level: SAFEPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill is designed to analyze up to 2,000 files from external data rooms or cloud storage platforms (e.g., Feishu, Nutstore). This represents a significant ingestion surface for indirect prompt injection, where malicious instructions could be embedded in contracts or legal documents to subvert the agent's logic or extract information.\n
  • Ingestion points: Processes external documents from various sources (MCP tools, cloud storage) as well as matter-specific local files.\n
  • Boundary markers: Employs source attribution tags like [用户提供] and [数据室路径] to categorize data, though these do not serve as hard security boundaries against instruction following.\n
  • Capability inventory: File system read/write access to the user's home directory (~/.claude/plugins/config/) and integration with third-party legal research tools.\n
  • Sanitization: No explicit sanitization or filtering of incoming document content for prompt injection patterns is described.\n- [DATA_EXFILTRATION]: The skill accesses sensitive local file paths within the home directory (~/.claude/plugins/config/claude-for-legal/) to load matter context, deal thresholds, and internal configurations. While these paths are application-specific, the ability to read and write to these locations constitutes a potential data exposure risk if the agent's context is manipulated.
Audit Metadata
Risk Level
SAFE
Analyzed
May 15, 2026, 07:06 AM
Security Audit — agent-trust-hub — diligence-issue-extraction