wechat-labcanvas-chatops

Fail

Audited by Gen Agent Trust Hub on Oct 19, 2026

Risk Level: HIGHDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPERSISTENCEDYNAMIC_EXECUTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill accesses highly sensitive private data, including encrypted WeChat message databases located at xwechat_files/<WXID>/msg/, private decryption keys (<PRIVATE_KEY_FILE>), and local configuration files (.env, .aginti/private/). While instructions are provided to keep these private, the agent itself has broad read access to these files to perform its functions.
  • [COMMAND_EXECUTION]: Employs a wide range of system commands and CLI tools, such as labcanvas, xdotool, Xvfb, x11vnc, websockify, ffmpeg, and git. It automates a headless GUI and executes various tasks that involve direct system interaction and automation of the official WeChat client.
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from WeChat messages and uses it to drive agent behavior and routine selection.
  • Ingestion points: Reads messages and file metadata from local SQLite databases (msg/, file/) and synced media directories.
  • Boundary markers: Uses profile-specific state paths and implements a "danger policy" instructing the agent to return NO_REPLY for requests involving secrets or bot rule changes.
  • Capability inventory: Includes sending files/artifacts, GUI automation via xdotool, and a "Worker Agent" framework for tasks like CAD export, rendering, and media processing.
  • Sanitization: Instructs not to pass raw aeskey or voiceurl XML into prompts, but lacks robust sanitization for general message content used in routing.
  • [PERSISTENCE]: Utilizes tmux sessions and restart wrappers to maintain persistent, long-running processes for monitoring WeChat message streams, managing worker task queues, and periodic media synchronization.
  • [DYNAMIC_EXECUTION]: Implements a multi-agent routing architecture where input from chat messages is routed to workers that dynamically execute "routines" and tool-specific reasoning based on the content of incoming WeChat messages, which can include complex workflows like CAD/PCB generation and media transcription.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 19, 2026, 06:24 PM