The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of NIH grant abstracts and titles. It includes specific mitigations in SKILL.md, instructing the agent to treat this metadata as evidence only and to ignore any embedded commands. The evals/evals.json file contains test cases (ID 10) specifically designed to verify that the agent resists instructions hidden within grant abstracts.
Ingestion points: Data enters through OpenNIH_search_grants, OpenNIH_fetch, and NIHReporter_search_projects.
Boundary markers: The skill uses explicit instructional boundaries to separate grant content from agent instructions.
Capability inventory: Capabilities are limited to data retrieval, synthesis, and cross-referencing with scientific databases; no arbitrary file-write or shell execution tools are enabled for user data processing.
Sanitization: The instructions mandate treating all retrieved metadata as untrusted evidence.
[DATA_EXPOSURE_EXFILTRATION]: The skill communicates with opennih.org (a vendor-controlled domain) and well-known public scientific services including PubMed, iCite, and ClinicalTrials.gov. No sensitive local credentials or system files are accessed.
[REMOTE_CODE_EXECUTION]: The provided Python scripts in the scripts/ directory are intended for local regression testing and contract verification of the API. These scripts interact only with defined scientific endpoints and do not download or execute code from untrusted remote sources.
[PROMPT_INJECTION]: The skill uses natural instructional language to enforce rigorous data handling. Analysis of the evaluation prompts shows a focus on preventing the agent from making unsupported claims (e.g., estimating application success odds) or being misled by data content.

tooluniverse-nih-funding-landscape