Skip to content

Security: kgraph57/paper-writer-skill

Security

SECURITY.md

Security Policy

Normal Skill Behavior

This package is a documentation-first Agent Skill for medical/scientific manuscript writing. Unlike zero-network skills, it intentionally uses network access for literature work: WebSearch/WebFetch and literature APIs (PubMed, OpenAlex, Europe PMC, Semantic Scholar) during Phase −1 novelty checks, Phase 1 literature search, and Phase 5 citation verification. No credentials are required for any of these.

The utility scripts under scripts/ are optional. They read local files, write local Markdown/plot outputs, and make no network calls. Analysis and PDF utilities use the Python packages listed in requirements.txt.

Clinical Data Handling — the hard rules

This skill is built for clinicians working with real patient data. The data rules below mirror templates/data-management.md and the AI-for-Science model (references/ai-for-science-model.md):

  • Raw clinical data stays local. Always. Generated project directories keep original data under data/raw/, which is treated as read-only and is gitignored — it must never be committed or pushed anywhere.
  • Never commit PHI. Only de-identified, processed data belongs in data/processed/. De-identification happens before anything leaves the raw folder.
  • Never paste identifiable patient information into any AI chat, including the agent running this skill. Work from de-identified extracts.
  • AI never originates data. The skill must never fabricate a data point, participant, or result (📊 DATA sovereignty). Any output that appears to contain invented data is a bug — report it.

Permissions

Capability Used Notes
Network access Yes Literature APIs and web search only; no telemetry
Filesystem write Yes Manuscript project directories and script outputs
External tools Optional Python packages from requirements.txt; pandoc for docx/pdf compile
Credentials No No API keys required
Background execution No

Safety Rules

  • Do not add hidden scripts or install-time automation.
  • Do not add new network destinations without documenting them here.
  • Do not embed real patient data — even "harmless" fragments — in examples, tests, or case studies. All repo examples are illustrative and PHI-free.
  • Do not weaken the 💡 IDEA / 📊 DATA human-sovereignty gates in SKILL.md.

Reporting

Open a GitHub issue for suspected vulnerabilities, unsafe instructions, or hidden behavior. Do not include patient data or other confidential material in public reports.

There aren't any published security advisories