The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is instructed to dynamically fetch and process model documentation and schemas from external sources such as docs.higgsfield.ai and console.higgsfield.ai. This creates a theoretical surface for indirect prompt injection if the remote content were to be compromised.
Ingestion points: The agent is guided to read llms.txt files and OpenAPI references from external URLs to discover model capabilities (found in SKILL.md and references/model-discovery.md).
Boundary markers: The skill provides explicit instructions to the agent that documentation should be treated as data, not as authorization to execute commands, and requires verifying documentation against local profiles.
Capability inventory: The skill possesses the capability to perform network operations (GET/POST/PUT), write files to the project's outputs/ and work/ directories, and read local files for media uploads.
Sanitization: The scripts/higgsfield_api.py script implements strict schema validation using the jsonschema library for all model profiles and payloads. It also enforces environment checks and requires a verification date for all profiles.
[EXTERNAL_DOWNLOADS]: The skill downloads generated media files from remote URLs provided by the API response.
Evidence: The download function in scripts/higgsfield_api.py uses urllib.request to retrieve content from external hosts.
Mitigation: The script includes robust SSRF (Server-Side Request Forgery) protections in the public_url function, which validates that hostnames resolve to global (public) IP addresses and prevents access to local or private network ranges. It also enforces a 256 MiB file size limit.
[DATA_EXFILTRATION]: The skill includes an upload capability that reads local media files and transmits them to external storage hosts.
Evidence: The upload function in scripts/higgsfield_api.py reads local file bytes and performs a PUT request to a presigned URL.
Mitigation: The script limits uploads to specific allowed media MIME types and a 64 MiB size limit. It also contains logic to ensure that internal API credentials are never sent to the third-party storage host.
[COMMAND_EXECUTION]: The skill operates by executing a local Python script (higgsfield_api.py) to manage the API lifecycle, job state, and file operations.

higgsfield-api