The Agent Skills Directory

[PROMPT_INJECTION]: No malicious injection patterns or safety filter bypass attempts were detected. The skill instructions contain extensive safety boundaries and negative constraints that prohibit the agent from providing medical advice, diagnosing, or treating patients.\n- [EXTERNAL_DOWNLOADS]: The skill references numerous official medical and regulatory domains (e.g., ich.org, hhs.gov, fda.gov) for research guidance. It also instructs the agent to fetch citation metadata from arXiv (https://arxiv.org/abs/2609.00065). These are well-known services relevant to the skill's scientific purpose and do not represent a security risk.\n- [DATA_EXFILTRATION]: No exfiltration patterns detected. The accompanying Python scripts utilize a shared validation module (_common.py) that enforces strict local-only file access, blocks URL/URI schemes, and prevents the use of symbolic links. The skill is designed to work exclusively with de-identified or aggregate data.\n- [INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing external JSON and CSV clinical manifests.\n
Ingestion points: Local data files are ingested by scripts such as validate_case_report.py and format_adverse_events.py.\n
Boundary markers: The skill uses strict JSON schemas and template assets to define data boundaries.\n
Capability inventory: Capabilities are limited to deterministic formatting and consistency checking; no system commands or arbitrary code execution are performed on the ingested data.\n
Sanitization: Input fields are validated against regular expressions for identifiers and specific clinical terminology formats.\n- [DYNAMIC_EXECUTION]: No dangerous dynamic execution patterns were identified. The scripts perform structural validation and arithmetic checks using standard library modules. generate_report_template.py safely copies local assets to user-specified paths with validation.

clinical-reports