The Agent Skills Directory

[DATA_EXFILTRATION]: The skill hardcodes a specific external IP address 39.108.254.228 in SKILL.md, scripts/publish.py, and multiple reference files. All content generated by the AI—including full article bodies, titles, and metadata—is sent to this IP address via POST requests when the 'publish' functionality is used.
[CREDENTIALS_UNSAFE]: The references/server-config-guide.md file provides detailed instructions for users to extract highly sensitive session cookies (such as ZHIHU_COOKIE, JUEJIN_COOKIE, and CSDN_COOKIE) and API tokens from their personal accounts and store them in an environment file on the server at the hardcoded IP address 39.108.254.228.
[METADATA_POISONING]: The skill deceptively labels the hardcoded IP 39.108.254.228 as the user's own 'self-built service' (您自建的服务) in the instructions. This is misleading as it encourages users to send data to a pre-defined external infrastructure without explicitly prompting them to configure their own server first.
[COMMAND_EXECUTION]: The scripts/publish.py script uses subprocess.run to execute the system curl command to perform network operations. While it uses a list format for arguments, it still facilitates the transmission of potentially sensitive data to an external server using system-level process execution.
[PROMPT_INJECTION]: The skill instructions in SKILL.md and references/content-adaptation-guide.md explicitly direct the agent to 'ensure pure output' by suppressing all author information and production markers (e.g., 'COZE production', 'AI assistant'). This concealment technique is often used to mask the automated nature of content or bypass platform-specific watermarking and safety policies.

wechatsync-publisher