The Agent Skills Directory

[COMMAND_EXECUTION]: The authentication script (scripts/skywork_auth.py) utilizes subprocess.Popen with shell=True to open the login URL on Windows platforms. This pattern incorporates the SKYWORK_WEB_BASE environment variable into a shell-executed command string, which is a common practice for opening URLs but presents a potential injection vector if the environment is not controlled.
[DATA_EXFILTRATION]: The skill transmits local image files and user prompts to the official vendor API (api-tools.skywork.ai) for processing. This network communication targets the vendor's official infrastructure and is necessary for the core image generation and editing tasks.
[PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection by accepting untrusted user prompts and processing local images for use with a backend AI model.
Ingestion points: User-provided command-line arguments and local image files processed in scripts/generate_image.py.
Boundary markers: Not present in the script logic to delimit user input or warn the model about embedded instructions.
Capability inventory: Network communication with the vendor API and local file system access for reading images and storing authentication tokens.
Sanitization: No specific filtering or validation of the prompt text is performed before it is sent to the backend.

Skywork Design