The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script assets/daily_report.py attempts to read sensitive shell configuration files, including ~/.zshrc, ~/.bashrc, ~/.bash_profile, and ~/.profile, to extract the REDFOX_API_KEY. Reading these files exposes the entire contents of the user's environment configuration to the agent process, which is a high-risk data exposure pattern.- [PERSISTENCE_MECHANISMS]: The skill implements automated persistence by programmatically scheduling recurring tasks. Using the --subscribe flag, it installs a LaunchAgent on macOS (~/Library/LaunchAgents/com.qoder.bili-ai-feed.plist) or modifies the system crontab on Linux and Windows. This ensures the skill's code continues to execute periodically across user sessions.- [DYNAMIC_EXECUTION]: The script uses subprocess.run with shell=True to execute shell commands during the persistence installation process, specifically when piping data into crontab. This pattern is susceptible to command injection if local paths or identifiers were to be manipulated.- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill transmits search keywords and request metadata to the vendor's API at https://redfox.hk. While this is part of the core functionality, it involves sending potentially sensitive user-defined keywords to a remote server.- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from Bilibili video titles and descriptions. This content is ingested into the agent's context to generate reports and cluster topics. Without explicit boundary markers or sanitization, malicious content embedded in video metadata could influence the agent's output or subsequent tool invocations.

bili-ai-feed