The Agent Skills Directory

[COMMAND_EXECUTION]: The instructions in SKILL.md direct the AI agent to perform system-level modifications by appending commands to shell configuration files (e.g., ~/.zshrc, ~/.bashrc) or modifying Windows environment variables at the user level to persist API keys.
[DATA_EXFILTRATION]: The script scripts/douyin_similar_account.py contains functions (_read_api_key_from_unix_shell_config and _read_api_key_from_windows) designed to read sensitive local files, including ~/.zshrc, ~/.bashrc, ~/.bash_profile, and ~/.profile, searching for configuration strings and credentials.
[PROMPT_INJECTION]: The skill includes explicit instructions to the AI agent to ignore its standard operational procedures. Specifically, SKILL.md and references/core_workflow.md mandate that the agent must display script output "as-is" and "原样完整展示" (completely as originally presented), forbidding any summary, modification, or additional commentary, which suppresses the agent's ability to filter or review content.
[DYNAMIC_EXECUTION]: The wrapper script query_wrapper.py utilizes the exec() function to run the contents of the core logic script. Dynamic execution of file contents can be used to bypass security scanners and presents a risk of executing unauthorized code if the underlying file is modified.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow where untrusted data fetched from a remote API (redfox.hk) is fed directly into the agent's response context. Combined with the instruction to bypass summarization, this creates a vulnerability where malicious instructions returned by the API could be executed by the agent or directed at the user without oversight.

douyin-similar-account