Switchboard handles credentials and controls what AI coding agents can reach, so security reports get priority over everything else.

Please do not open a public issue for anything security-sensitive.

Use GitHub's private vulnerability reporting: Report a vulnerability.

You can expect an acknowledgment within a few days. Switchboard is a solo-run alpha project, so there is no security team or bounty program, but real reports get fixed fast and credited if you want credit.

The full security analysis, including trust boundaries and the accepted-risks list, is in the threat model. If shipped behavior contradicts that document, the contradiction itself is in scope.

Reports that matter most:

• Secret values surfacing anywhere: output, logs, JSON, audit entries, MCP responses, backups Switchboard itself creates.
• A pass failing to enforce its policy on a routed path: denied tools reachable, expired passes still working, approval gates bypassed through Switchboard.
• Keychain handling flaws, including unsafe fallback activation without the explicit opt-in.

Known and documented limits, not vulnerabilities:

• Switchboard only governs paths routed through it. Raw shell access, direct MCP routes, provider CLIs, and browser sessions can bypass it. switchboard scan exists to surface those routes; the README says this plainly.
• Rollback backups are exact copies of prior config and may contain tokens that were already in plaintext before import.

Alpha: only the latest published version of @switchboard-mcp/cli is supported. There are no backports.