The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data to generate clinical report drafts, which presents an attack surface for indirect prompt injection. However, the skill is heavily mitigated by its design.
Ingestion points: Untrusted data enters the context through user-provided JSON manifests and CSV aggregate files processed by the scripts in the scripts/ directory.
Boundary markers: The skill enforces strict boundary markers by prepending mandatory safety notices and draft warnings to all outputs, and it uses fail-closed statuses like BLOCKED_INCOMPLETE to prevent accidental use.
Capability inventory: The skill is limited to local file system reading and writing of specific clinical schemas; it contains no network access, subprocess execution, or dynamic code evaluation capabilities.
Sanitization: Input validation is handled in scripts/_common.py, which enforces file size limits, JSON depth limits, and string length restrictions, and explicitly rejects non-local paths and URI schemes.
[SAFE]: The skill implements robust local-only processing and adheres to medical data handling best practices.
Security Measures: The scripts/_common.py utility ensures that all file operations are bounded to the local filesystem and rejects any input containing URI schemes (e.g., http://, file://), preventing remote resource access.
Trusted Sources: All external URLs included in the documentation point to well-known official regulatory bodies and scientific organizations such as the FDA, HHS, ICH, and established medical journals, qualifying as safe references.

clinical-reports