Skip to content

Security: itsmeadarsh2008/opencode-senses

Security

SECURITY.md

Security Policy

OpenCode Senses is a local-first vision plugin. Images and analysis never leave your machine unless you explicitly opt into a remote provider (senses_reverse with yandex/saucenao/tracemoe, or Moondream hosted finetunes via MOONDREAM_API_KEY).

Supported versions

Version Supported
Latest npm release Yes
main branch Yes (best effort)
Older releases No

Reporting a vulnerability

Please do not open a public issue for security problems. Report privately instead:

  • Email: reach out via a private security advisory
  • GitHub: use the "Report a vulnerability" button on the repository's Security tab

Please include, where possible:

  1. A description of the vulnerability and its impact
  2. Steps to reproduce (minimal, please)
  3. Affected versions and configurations
  4. Suggested fix, if you have one

You will receive a response within 7 days, and we will work with you on a coordinated disclosure. If you found a real issue, you also have our genuine thanks and a place in the release notes (if you want one).

Security-relevant areas

Things we consider in-scope for security review:

  • Prompt injection: any new code path that lets image-derived text reach the model outside the <SENSES> untrusted-data guards
  • Remote fetch: path accepting http(s):// URLs — SSRF considerations, size/timeout handling, file-type verification
  • Path handling: materialized files in /tmp, crops/zooms/annotations in the cache dir, reverse-search index — no path traversal, no symlink tricks
  • Runtime provisioning: the auto-installed Python venv and model weight downloads are checksum-less third-party artifacts; treat supply-chain risk seriously
  • Local-only guarantee: no hidden telemetry, no background uploads, ever

Non-issues

  • Bugs, crashes, or poor OCR quality are normal issues, not vulnerabilities. Open them as regular issues.
  • The vision model being imperfect at reading text is a feature request, not a breach. It is, after all, a small model with big feelings.

There aren't any published security advisories