Please report suspected security issues privately — do not open a public GitHub issue.
Open a GitHub Security Advisory in this repository. That's the preferred channel; it's private by default and lets us coordinate a fix before disclosure.
You should get an acknowledgement within 7 days.
Rouge is an autonomous product-development system that runs Claude Code with
--dangerously-skip-permissions and manages real cloud infrastructure
(Vercel, Cloudflare, Supabase, GitHub). The threat surface includes:
- The launcher (
src/launcher/) — safety hooks, deploy blocking, secrets handling. - Safety hooks (
src/launcher/rouge-safety-check.sh) — the PreToolUse/PostToolUse hook that blocks dangerous commands. - Secrets backend (
src/launcher/secrets.js) — cross-platform credential store integration. - Slack bot (
src/slack/) — if you run the Slack control plane. - Dashboard (
dashboard/) — the Next.js control plane, its API routes, and SSE transport. - Integration catalogue (
library/integrations/) — the patterns Rouge draws from.
- Products Rouge builds for you. Their security is their own concern — they get their own repos, deploy targets, secret scopes.
- Claude Code itself. Report those to Anthropic.
--dangerously-skip-permissions. This is a deliberate product choice, documented in the README. The safety hooks + deploy blocking are the defence, not the permission prompt.- The host machine. Rouge assumes it runs on a dedicated machine or VM. Data exposure caused by running Rouge on a shared box is not a Rouge vulnerability.
Rouge is not a hardened multi-tenant system. It's designed for a single developer (or small team) running it against their own Claude subscription and cloud accounts.
Reporters who follow responsible disclosure are credited in the release notes of the version that contains the fix, unless they request otherwise.