Skip to content

Security: BuildHarbor/sciagent-skills

Security

SECURITY.md

Security Policy

SciAgent Skills is a library of plain-text SKILL.md instructions. The skills themselves are not executable code and run with whatever privileges your coding assistant already has. The security considerations are therefore about what the skills tell the agent to do, not about the repo running anything itself.

Supported versions

Version Supported
1.x yes

We patch the latest main and cut patch releases for security-relevant fixes.

Reporting a vulnerability

Do not open a public issue for security reports.

Email: security@<Pratikrishi97>.dev (replace with your real address) with the subject sciagent-skills security report. You will receive an acknowledgement within 72 hours. We coordinate disclosure and credit reporters in the release notes unless they prefer to remain anonymous.

Please include:

  • A description of the issue and its impact.
  • A minimal reproduction (a skill snippet, a prompt, or a config).
  • Affected versions / commits.
  • Any suggested fix.

Threat model

  1. Prompt injection via skill content. A malicious skill could instruct the agent to exfiltrate data or run destructive commands. Mitigation: only install skills from sources you trust. Review any third-party skill's SKILL.md before installing. The allowed-tools experimental frontmatter field lets a skill pre-approve a limited tool surface — prefer skills that use it.
  2. Secrets in skills. Skills are plain text in a public repo. Rule: never put API keys, tokens, credentials, or real patient identifiers in a skill. Snippets should use placeholders (<YOUR_API_KEY>), os.environ.get(...), or local-only config.
  3. Clinical data and PHI. Skills in the Clinical & Medical Research category assume data is already de-identified or synthetic. Rule: never paste real PHI into a prompt. Use synthetic datasets (MIMIC-IV is already de-identified but still treat with care) and assume any text you send to a hosted assistant may be logged.
  4. Dual-use research. Synthesis, retrosynthesis, and pathogen-genomics skills are for legitimate, open research. Rule: refuse requests whose only plausible use is weapons, and refer users to their institutional biosafety committee (IBC) / dual-use research of concern (DURC) review process.
  5. Lab safety. Chemistry and biology skills describe in silico workflows and literature-scale reasoning. Rule: they are not laboratory SOPs. Wet-lab execution requires institutional review, PPE, and proper training — none of which a skill can substitute for.
  6. Supply chain. Skills may recommend pip install lines. Mitigation: only recommend packages from PyPI with established maintainership; prefer conda-forge for compiled scientific libraries. The repo's own dependencies (pyproject.toml) are pinned and scanned by CI.

Hardening checklist for skill authors

  • description includes trigger keywords but no secrets.
  • compatibility lists the libraries the skill needs; nothing exotic.
  • Snippets use os.environ.get() for any credential, never a literal.
  • Clinical/genomics snippets operate on de-identified or synthetic data.
  • Ethics & safety section is present and specific (not boilerplate).
  • No curl | bash, no eval, no exec of remote content.
  • allowed-tools (if used) is the minimal set the skill actually needs.

Incident response

  1. Confirm the scope (which skill, which version, which agents).
  2. Mark the affected skill folder with a SECURITY-NOTE.md describing the issue and workaround.
  3. Cut a patch release; update CHANGELOG.md with a security entry.
  4. Publish a GitHub Security Advisory and request a CVE if warranted.
  5. Notify known downstream consumers (the plugin.json marketplace, if listed).

There aren't any published security advisories