The Agent Skills Directory

[COMMAND_EXECUTION]: The skill operates via a Python script (scripts/generate.py) that performs various system-level operations including creating directories, reading local configuration files (.env), and writing generated content (images, JSON manifests, and HTML files) to the local filesystem.
[DATA_EXFILTRATION]: The script is designed to transmit the AZURE_OPENAI_API_KEY to the specified AZURE_OPENAI_ENDPOINT. While this is necessary for functionality, it presents a risk if the endpoint environment variable is manipulated to point to a malicious server, leading to credential harvesting.
[PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect injection and content poisoning in its HTML gallery generation logic.
Ingestion points: The script receives untrusted data in the form of the user-provided --prompt and the revised_prompt field returned from the Azure OpenAI API in scripts/generate.py.
Boundary markers: None present; the script does not use delimiters or warnings to isolate the prompt content from the HTML structure.
Capability inventory: The script has permissions to write files (open(), Path.mkdir()) and perform network operations (urllib.request.urlopen()).
Sanitization: Absent. The create_gallery_html function in scripts/generate.py performs direct string interpolation of prompts into the HTML body without escaping special characters (e.g., < > " ' &), which could allow for the execution of malicious scripts when the gallery is viewed in a browser.

azure-image-gen