Skip to content

Security: indranilbanerjee/cutforge

Security

SECURITY.md

Security Policy

Supported Versions

The latest minor release of CutForge receives security fixes. Please update before reporting.

Version Supported
1.1.x ✅
1.0.x ⚠️ Security fixes only when trivially backportable
< 1.0 ❌ Please upgrade

Reporting a Vulnerability

Do NOT open a public GitHub Issue for security vulnerabilities. Public issues expose users before a fix exists.

Use GitHub's Private Security Advisory to report privately. If you cannot, contact the maintainer through the email listed on indranil.in with the subject line [CutForge Security].

Please include: a description and its impact, steps to reproduce, the affected version and surface (Claude Code, Cowork, Codex, Cursor, Copilot CLI, Antigravity, Hermes Agent, OpenClaw or Grok), a suggested fix if you have one, and whether you would like credit.

What you can expect: acknowledgement within 48 hours, an initial assessment within 7 days, coordinated disclosure, and credit in the advisory unless you prefer otherwise. There is no bug bounty: this is an open-source project maintained by an individual, and responsible disclosure is deeply appreciated.

Scope

In scope:

  • The plugin itself: skills, agents, manifests and the Hermes adapter
  • The engine in engine/ (any path traversal, command injection or unsafe deserialization in the Python, shell or Node tools, including what a scaffolded project runs)
  • The quote gate (tools/spend.py, tools/price_book.py) and the face guard (tools/vfx/ai_gen.py): a way to bypass either is in scope

Out of scope:

  • Vulnerabilities in Claude Code, Cowork or other hosts (report to their vendors)
  • Vulnerabilities in Remotion, ffmpeg, Chromium or upstream Python packages (report upstream, and tell us so we can pin a fixed version)
  • Third-party provider APIs (Sarvam, ElevenLabs, OpenAI, WaveSpeed)
  • Reports from automated scanners without proof of exploitability

Hardening recommendations for operators

  1. Keep API keys out of projects you share. Tools read keys from environment variables or the key file named in a project's config.json; keep that file outside the project folder and never commit it.
  2. Treat client material as personal data. Projects hold family photos, children's faces, voices and documents. Apply the folder permissions you would to any client PII, and delete working copies when the job ends.
  3. Keep the quote gate on. Run a paid tool with --go only after the client approves its quote, and keep budget_inr set.
  4. Review skill and tool changes in PRs like production code: skills run with whatever permissions your host has.
  5. Pin the plugin version for client work and update between jobs, not during one.

There aren't any published security advisories