The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is instructed to ingest and process data from untrusted external sources to make deployment and troubleshooting decisions.
Ingestion points: The agent reads ComfyUI workflow.json files, HuggingFace/Civitai repository metadata, and system/worker logs (stream-worker-logs) to determine hardware requirements, engine selection, and failure causes.
Boundary markers: The instructions do not define specific delimiters or instructions for the agent to ignore potentially malicious embedded commands within these external files.
Capability inventory: The skill has broad capabilities, including creating and deleting infrastructure (create-endpoint, create-network-volume) and triggering job execution (run-endpoint).
Sanitization: The skill lacks specific sanitization logic for the content of processed workflows or log output beyond basic string matching for file extensions.
[EXTERNAL_DOWNLOADS]: The skill facilitates the deployment of community-authored (non-vendor) templates and models from external hubs.
Mitigation: The skill includes a mitigation by requiring the agent to 'name the repo and its owner from the listing, never from memory' and instructs against inventing URLs. It also utilizes official vendor-managed tools (list-hub-repos, deploy-hub-repo) for these operations.

runpod-serverless-deploy