cc-haha

Least-privilege permissions

• Restrict agent access: Limit what actions agents can perform on your computer. Only grant "Computer Use" (ability to control your screen/keyboard) to trusted agents; disable it for experimental or third-party tools.
• Isolate messaging account credentials: Store WeChat, Feishu, DingTalk, and other login tokens in a separate, encrypted vault. Never hardcode or share these credentials; rotate them monthly.
• Sandbox skills from the marketplace: Treat downloaded agent skills like app downloads—only install from verified publishers. Use a sandboxed test environment before running them on your main workspace.
• Separate user accounts for testing: Run new or untrusted agents under a limited computer user account (not your admin account) so they cannot access sensitive files or system settings.

Network isolation

• VPN or firewall rules for remote access: If using WeChat/Feishu/DingTalk remote control, restrict access to a corporate VPN or whitelist specific IP ranges. Do not expose remote endpoints to the public internet.
• Encrypt all messaging connections: Ensure all remote messaging platforms enforce end-to-end encryption. Verify that tokens and session data are transmitted over HTTPS only.
• Monitor outbound traffic: Use a network tool (like your firewall or packet sniffer) to confirm no unexpected data is leaving your computer to external servers. Especially check for telemetry or logs going to unmapped cloud services.
• Disable remote access when not needed: Turn off WeChat/Telegram/H5 remote access during sensitive work. Only enable the minimum messaging platforms required for your workflow.

Monitoring

• Review agent action logs weekly: Check what actions each agent performed (files accessed, commands run, data copied). Most MCP tools log this—look for unusual patterns like access to password managers or payment files.
• Audit skill installations and updates: Keep a list of installed marketplace skills and their last update date. Investigate any auto-updates or new versions that perform new actions.
• Track failed authentication attempts: Monitor failed login attempts on your messaging accounts. A spike may indicate someone is trying to gain remote access to your workspace.
• Set alerts for sensitive file access: Configure your OS to alert you if agents access critical folders (passwords, financial data, source code repositories). Flag any unexpected access immediately.

Red flags to watch for

• Agents requesting "full system access" or refusing permission scoping: Legitimate agents should accept restrictions. Any tool demanding unrestricted Computer Use privileges is suspicious.
• Skills updating themselves automatically without your approval: Third-party skills should not auto-update. Any self-modifying behavior suggests potential malware or supply-chain compromise.
• Unexpected network traffic or battery drain: If your computer is slow or battery depletes quickly, agents may be running hidden processes or exfiltrating data in the background.
• Credential or token prompts appearing unexpectedly: If the workspace suddenly asks for your WeChat password, API keys, or bank login, do not enter it—this may be a prompt injection attack from a compromised skill or agent.