The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The toolkit utilizes standard libraries such as Pillow and NumPy and executes external processes via subprocess. Evidence: scripts/long_screenshot_ocr.py and scripts/html_shot.py use argument lists for process invocation, mitigating command injection risks.
[DATA_EXPOSURE_AND_EXFILTRATION]: Vision API credentials are used exclusively for their intended purpose. Network activity is limited to localhost for browser control. Evidence: scripts/html_shot.py logic is confined to 127.0.0.1 communication.
[INDIRECT_PROMPT_INJECTION]: The skill processes text from images, creating an attack surface for indirect instructions.
Ingestion points: scripts/long_screenshot_ocr.py extracts text from image chunks using an OCR utility.
Boundary markers: Instructions in ocr_prompt define formatting but do not include security delimiters for the extracted content.
Capability inventory: The skill can launch local browsers and perform file operations.
Sanitization: Transcribed text is not filtered for potential injection patterns before reaching the agent context.
[DYNAMIC_EXECUTION]: scripts/html_shot.py evaluates JavaScript within a headless browser to manage page layout and capture screenshots. Evidence: Use of Runtime.evaluate in the browser DevTools Protocol to query document state.

vision-skills