read-personal-data-harvester

Warn

Audited by Gen Agent Trust Hub on Jul 11, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill accesses highly sensitive local data, including the internal SQLite databases of the WeChat Read application ("~/Library/Containers/...") and browser profiles. This allows the agent to harvest extensive personal history and potentially access authenticated sessions to collect data from platforms like Douban and Bilibili.
  • [COMMAND_EXECUTION]: The skill establishes persistence on the host system by instructing the agent to modify the user's crontab and create a launchd agent ("~/Library/LaunchAgents/com.personal-harvest.plist"). This ensures the harvesting scripts run automatically in the background at scheduled intervals.
  • [PROMPT_INJECTION]: The skill processes untrusted data (book titles, comments, and post descriptions) from third-party websites. This data is interpolated into local database queries and agent context without sanitization, creating an attack surface for indirect prompt injection where malicious content on a webpage could influence agent behavior.
  • Ingestion points: External platform data (Douban, Bilibili, Xiaohongshu) ingested via Playwright and local file parsing.
  • Boundary markers: Absent. Content is directly mapped to dictionary fields and SQL queries.
  • Capability inventory: Subprocess calls (bash), file-write (SQLite), network operations (Playwright, requests).
  • Sanitization: No escaping or validation is performed on the scraped content before storage or processing.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jul 11, 2026, 03:50 PM
Security Audit — agent-trust-hub — read-personal-data-harvester