Please do not open a public issue for exploitable vulnerabilities.

Report privately through GitHub's private vulnerability reporting: github.com/abudoxali/RootRay/security/advisories/new

Include: affected version, reproduction steps, impact, and any suggested mitigation. We will acknowledge the report and coordinate a fix and disclosure timeline with you. Do not publicly disclose details before a fix is available.

RootRay is local-first by design — understanding the boundaries helps target reports usefully:

• No RootRay-owned remote listener. The inspector bridge binds to ws://127.0.0.1 (loopback only) with an ephemeral per-session token on a dynamic port. Wrong token/session/version and malformed messages are rejected. A launched project dev server retains its own network policy.
• Preview isolation. The embedded project Preview is a separate WebView2 with zero IPC privileges — every privileged Tauri command is denied to it natively. Main-frame navigation is loopback-only; window.open/target=_blank is denied; remote URLs open in the system browser unprivileged.
• Filesystem boundary. All file access is relative to the selected workspace root: .. traversal, absolute paths, symlink escapes, .env*, keys/credentials, .git, node_modules, target, dist, binary and oversized files are refused.
• Safe writes. Saves use SHA-256 optimistic concurrency and same-directory atomic rename; external modifications surface as conflicts rather than silent clobbering.
• Process containment. Dev servers run inside a Windows Job Object with KILL_ON_JOB_CLOSE; unrelated processes are never touched and nothing is killed by port.
• Narrow command surface. The UI can only invoke an audited set of Tauri commands — there is no generic command API.
• Discovery is inert. Workspace analysis never executes project code, never reads secrets, and is bounded (depth/count/size caps).

• Vulnerabilities in the user's own dev server or application — RootRay runs your project but doesn't proxy it.
• Issues requiring an already-compromised local account (the security boundary assumes the Windows user account is trusted).