The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill documentation in SKILL.md and associated launcher scripts (launch_local.ps1) instruct the user to execute PowerShell with the -ExecutionPolicy Bypass flag. This circumvents local security restrictions designed to prevent the execution of untrusted scripts.
[DYNAMIC_EXECUTION]: The script scripts/process_chunks_local.py uses importlib to dynamically load whisper_dispatcher.py from a path determined at runtime. This practice can be used to execute arbitrary code if the target path is compromised.
[COMMAND_EXECUTION]: Multiple scripts, including scripts/_vision.py, scripts/anibon/lmstudio.py, and scripts/anibon/ytdlp.py, use subprocess.run to invoke external binaries. In lmstudio.py, the command is executed with shell=True. These commands use arguments, such as URLs and file paths, that are influenced by runtime variables.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from YouTube automatic transcripts and live chat messages (e.g., in scripts/process_chunks_local.py). This content is interpolated into LLM prompts without rigorous sanitization, allowing for potential prompt injection attacks embedded in the video content or chat history.
[EXTERNAL_DOWNLOADS]: The pipeline includes scripts (scripts/fetch_fgo_db.py, scripts/fetch_ygo_db.py) that fetch data from community gaming APIs (api.atlasacademy.io, db.ygoprodeck.com) to build local reference databases. It also uses an automated search module (scripts/anibon/websearch.py) to query DuckDuckGo Lite.

anibon-timestamper-local