The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Automated security scanners identified references/environment.md as malicious (FileRepMalware), which indicates the presence of a known threat or suspicious content within the skill's documentation files.\n- [PRIVILEGE_ESCALATION]: The skill contains explicit instructions in SKILL.md and references/environment.md that direct the user or agent to bypass security sandboxes (re-run with the sandbox disabled) if execution is interrupted, which is a common technique for evading security constraints.\n- [EXTERNAL_DOWNLOADS]: The script scripts/build_reel.py uses urllib.request.urlretrieve to fetch font files from the internet. While the default points to a trusted Google repository, the URL is fully configurable via the style.font_url field in the user-provided spec.json, allowing the skill to download arbitrary files from any remote source.\n- [COMMAND_EXECUTION]: Multiple scripts (build_reel.py, cut_timeline.py, map-words.py, verify_cuts.py) use the subprocess module to execute ffmpeg, ffprobe, and whisper-cli commands. These commands are constructed using parameters and file paths derived directly from the untrusted spec.json configuration, creating a significant attack surface for command injection.\n- [INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its ingestion of external configuration and media data.\n
Ingestion points: Processes user-supplied spec.json and external video/audio source files.\n
Boundary markers: None identified; the skill does not use delimiters or provide instructions to ignore embedded commands in the processed data.\n
Capability inventory: Executes shell commands via subprocess and makes network requests via urllib.\n
Sanitization: No validation or sanitization of input fields from spec.json was found before they are used in critical operations such as file path construction or command execution.

reel-recut