Skip to content

SECURITY.md

Latest commit

 

History

History
56 lines (42 loc) · 2.61 KB

SECURITY.md

File metadata and controls

56 lines (42 loc) · 2.61 KB

Security Policy

dexta-intelligence handles personal health data and ships connectors that authenticate to third-party services. We take security and privacy reports seriously.

Supported versions

The project is alpha (0.1.0). Only the main branch is supported. There is no published PyPI release yet; fixes land on main.

Reporting a vulnerability

Do not open a public issue for a security or privacy vulnerability.

Report privately through GitHub's private vulnerability reporting ("Security" tab → "Report a vulnerability"). If that is unavailable, open a minimal public issue that says only "security report - please enable private contact" without details, and a maintainer will follow up.

Please include, where you can:

  • the component (connector, store, guard, server/GUI, CLI, eval);
  • a description of the issue and its impact;
  • steps to reproduce or a proof of concept;
  • affected version / commit;
  • any suggested remediation.

We aim to acknowledge a report within a few days and to keep you updated on remediation. We will credit reporters who want it once a fix is available.

Scope and what to look for

The highest-value areas:

  • Secrets handling. Credentials live in environment variables, never committed config. Report anything that logs, persists, transmits, or renders a secret. The settings panel must only show set/unset status, never values.
  • Data egress. Dexta is self-hosted and your data stays on infrastructure you control. Report any unexpected network call, telemetry, or phone-home. Note that you opt into data egress when you configure a hosted LLM provider (prompts include computed evidence) or an unofficial connector (credentials go to the vendor endpoint) - see MEDICAL_DISCLAIMER.md and the connector tiers in docs/CONNECTORS.md.
  • The safety guards. guard/faithfulness.audit (prose may not cite numbers absent from evidence) and guard/treatment_gate (no dosing/treatment output) are safety-critical. A way to make the system emit a fabricated figure or dosing/treatment advice is a security-class bug - report it here, not as a normal issue.
  • Connectors. Unofficial / reverse-engineered connectors talk to vendor APIs. Report credential-handling or injection issues.

Out of scope

  • Vulnerabilities in third-party providers (Nightscout, Dexcom, LibreLinkUp, hosted LLM vendors) themselves - report those upstream.
  • The inherent risk of reverse-engineered connectors breaking when a vendor changes their site. That is a documented maintenance reality, not a vulnerability.