The Agent Skills Directory

[COMMAND_EXECUTION]: Several scripts, including benchmark_check.py, detect_hallucinations.py, and ytdlp.py, use subprocess.run to execute external tools like yt-dlp and whisper-cli. These commands take parameters like URLs and file paths that, if maliciously crafted, could lead to command injection or unintended system operations.
[DYNAMIC_EXECUTION]: The detect_hallucinations.py script computes a path to an external script in the user's home directory (~/.agents/skills/whisper-corruption-recovery/scripts/fix_hallucinations.py) and executes it using the current Python interpreter. This cross-skill execution from a computed path increases the risk of executing untrusted code if the target directory is compromised.
[DYNAMIC_EXECUTION]: The scripts/anibon/markdown.py file uses the __import__ function to dynamically load the sys module. While used here for logging, this technique is a common indicator of attempts to obscure dependencies or bypass static analysis.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcript data from YouTube and classifies it using a list of keywords in analyzer.py. This classification includes highly sensitive political terms, creating a vulnerability where an attacker could influence the agent's context or behavior by including specific triggers in a video's audio or captions (Mandatory Evidence: Ingestion in ytdlp.py, no boundary markers, capability for subprocess execution in detect_hallucinations.py, no sanitization).
[DATA_EXFILTRATION]: The SKILL.md file contains hardcoded Windows paths (e.g., C:\Users\peter\...) that expose specific local usernames and system directory structures.

anibon-local-transcription