The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The script scripts/fetch.py extracts sensitive authentication tokens (sessionid, csrftoken, ds_user_id) from local web browser databases. This logic targets Chrome, Safari, Firefox, Edge, and Brave to facilitate authenticated scraping of Instagram profiles.
[DYNAMIC_EXECUTION]: scripts/fetch.py defines a block of Python code in the IG_COOKIE_CODE string literal and executes it using subprocess.run. Additionally, scripts/look.py performs dynamic library loading via importlib.import_module to initialize OCR engines at runtime.
[COMMAND_EXECUTION]: Multiple scripts in the skill (including visual.py, sound.py, transcribe.py, and graphics.py) use subprocess.run or subprocess.Popen to execute system binaries such as ffmpeg, ffprobe, and yt-dlp for media processing and data collection.
[EXTERNAL_DOWNLOADS]: The skill downloads binary data from external sources, including a pre-trained face detection model (YuNet) from GitHub and account profile pictures from social media domains.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata and audio transcriptions from social media platforms, which creates a potential surface for indirect prompt injection attacks.
Ingestion points: Video metadata retrieved in fetch.py and transcripts generated in transcribe.py.
Boundary markers: Not present; untrusted data is interpolated into the agent context without clear delimiters.
Capability inventory: The skill can execute arbitrary shell commands via ffmpeg and yt-dlp, write analysis files to the local disk, and perform network requests.
Sanitization: No significant sanitization or filtering of external content is performed before processing.

creator-teardown