The Agent Skills Directory

[DATA_EXFILTRATION]: The core script scripts/check_sensitive_words.py reads sensitive shell configuration files, including ~/.zshrc, ~/.bashrc, ~/.bash_profile, ~/.profile, and ~/.zprofile, to search for the REDFOX_API_KEY credential. Accessing these personal configuration files without explicit per-file user consent is a security risk.
[DATA_EXFILTRATION]: User content, including direct text input, file contents, and data from provided URLs, is sent to an external API endpoint (https://redfox.hk/story/api/cozeSkill/sensitiveWordSearch) for analysis. This process transmits potentially private data to a third-party service.
[EXTERNAL_DOWNLOADS]: The skill requires the execution of playwright install chromium, which downloads and installs the Chromium browser binary from external sources to support its web scraping functionality.
[COMMAND_EXECUTION]: The agent workflow involves local file-system operations to generate and save optimized versions of user content as .txt files in the current directory.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it ingests and processes untrusted text from external documents and websites. This untrusted content is then used by the agent to generate output and perform file-system actions.
Ingestion points: The script extracts text from various user-provided sources, such as DOCX files, text files, and web links.
Boundary markers: There are no robust boundary markers or specific instructions implemented to isolate untrusted data from the agent's system instructions during processing.
Capability inventory: The agent has the capability to write to the local filesystem and interpret logic from the processed data.
Sanitization: The skill does not perform validation or sanitization of the content extracted from external sources before presenting it to the language model.

xiaohongshu-prohibited-word