The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface for indirect prompt injection by processing arbitrary local file paths and transmitting their contents to external AI model providers.
Ingestion points: The tools sn-image-recognize and sn-text-optimize in scripts/sn_agent_runner.py accept file paths via --user-prompt-path, --system-prompt-path, and --images arguments.
Boundary markers: The skill does not implement boundary markers or safety instructions to prevent the model from following instructions embedded in the processed data.
Capability inventory: The skill has the capability to read any file accessible to the agent process (via Path.read_bytes() in scripts/sn_image_base/vlm/utils.py and open().read() in scripts/sn_agent_runner.py) and perform network requests to external endpoints (via httpx in various adapter classes).
Sanitization: No sanitization of ingested content is performed before interpolation into API requests, allowing potentially malicious instructions in data to be processed by the backend models.

sn-image-base