The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes structured biomedical data and metadata from the external NCATS ARAX API, constituting a surface for indirect prompt injection.
Ingestion points: The scripts/arax_client.py script retrieves API responses from arax.transltr.io through /openapi.json, /entity, and /query endpoints.
Boundary markers: The script mediates interaction by parsing JSON responses and generating isolated artifacts (summary.json, response.json) and formatted text summaries, providing clear boundaries between raw external data and the agent's context.
Capability inventory: The skill writes query artifacts to a local directory and performs outbound HTTPS requests via urllib.request. It does not possess capabilities for arbitrary command execution or system modification.
Sanitization: Robust sanitization is implemented in scripts/arax_client.py via the _sanitize_text function, which removes control characters and truncates strings to 500 characters, effectively mitigating risks from malformed or malicious payloads.
[SAFE]: The skill demonstrates high-quality security practices including the following:
SSRF Protection: The validate_base_url function uses ipaddress.is_global to reject private, loopback, and reserved addresses, preventing Server-Side Request Forgery attacks.
Secure Redirects: A custom SameOriginHttpsRedirectHandler is implemented to reject cross-origin redirects and protocol downgrades (e.g., HTTPS to HTTP).
Secure File Handling: Artifacts are written using atomic operations and restricted file permissions (0o600), ensuring data integrity and minimizing exposure on the local filesystem.
Validated Network Operations: The skill communicates exclusively with the established biomedical API at arax.transltr.io over HTTPS using standard library components.

ncats-arax