The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external image data and user-provided prompts through the runcomfy CLI, which creates an attack surface for indirect prompt injection where malicious instructions could be embedded in the processed assets.
Ingestion points: Untrusted data enters the agent context via the --input JSON object in the runcomfy run command, which includes image, image_urls, and prompt fields (SKILL.md).
Boundary markers: The skill documentation contains a Security & Privacy section that instructs the agent to only ingest URLs explicitly provided by the user and warns to suspect reference assets if behavior diverges from the prompt (SKILL.md).
Capability inventory: The skill is restricted to the Bash(runcomfy *) tool, which is used to interact with the RunComfy Model API and write output to the local file system via the --output-dir flag (SKILL.md).
Sanitization: The documentation notes that the CLI uses JSON input to avoid shell-injection surfaces, but it does not specify any sanitization or filtering for instructions potentially embedded within the image data or prompts themselves.

relight