The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided text and image data, which are subsequently passed to external Vision Language Models (VLM) and Language Models (LLM). This creates a vulnerability surface for indirect prompt injection attacks.
Ingestion points: The --user-prompt and --images arguments in the sn-image-recognize and sn-text-optimize tools within sn_agent_runner.py.
Boundary markers: The skill does not implement delimiters or explicit instructions for the model to ignore embedded commands in the input data.
Capability inventory: The skill possesses the ability to make network requests to external APIs via httpx and write files (generated images) to the local filesystem.
Sanitization: The skill performs structural integrity validation on downloaded images using the Pillow library and implements a sanitization utility (sanitize_base64_in_data) to prevent large binary blobs from cluttering logs or error reports.
[EXTERNAL_DOWNLOADS]: The skill downloads generated image content from remote URLs provided by backend services.
Evidence: The download_image function in scripts/sn_image_base/generation/sensenova.py fetches data from URLs provided by the API response.
Context: The downloads originate from the CDNs of well-known and configured AI service providers (e.g., SenseNova, OpenAI, Anthropic), which are considered trusted sources in this context.

sn-image-base