Security

CrispyVibe is code-signed and notarized by Apple. All updates are verified with EdDSA signatures via Sparkle before installation.

Workspace integrity

• VibeSpace and project config files are HMAC-SHA256 signed
• If a config file is modified outside the app, CrispyVibe detects tampering on next load
• Tampered configs: the VibeSpace still opens (name, projects, colors visible) but startup commands are disabled
• A non-dismissable alert identifies the affected VibeSpace
• To restore: open VibeSpace Settings, review the configuration, and save (this re-signs the file)

Credential storage

• HMAC signing keys stored in macOS Keychain (scoped to app bundle ID)
• Authentication tokens (Cognito) stored in data-protection Keychain with afterFirstUnlockThisDeviceOnly accessibility
• SSH keys: key-based authentication only (no password auth); keys managed by system ssh-agent

SSH remote security

• Host key validation via ssh-keygen against known_hosts
• ControlMaster multiplexing with BatchMode=yes (no interactive prompts)
• File operations enforce project root boundary — attempts to escape return permission denied

Agent CLI security

• The control socket is owner-only (0600) — only processes running as the same local user can connect
• All file commands enforce project root boundary (including symlink resolution)
• Separate app instances use isolated sockets, preventing cross-instance access
• On remote (SSH) projects, the CLI is relayed over the SSH connection (per-connection opt-in); the forwarded socket is owner-only on the remote host

Terminal Insights

Terminal Insights validates screen visibility before recording input. When you type at a hidden prompt (sudo, ssh, read -s, GPG passphrase), the terminal disables echo — the typed text never appears in the render buffer and is never captured in command history or context summaries.

Reporting vulnerabilities

Email hello@crispyvibe.com. Do not open public issues for security vulnerabilities. Response within 48 hours.

🚨 Warning: Do not sign into sensitive accounts in browser panels shared with AI agents. Agents with browser access can read cookies and storage via eval.

Threat models

Each feature has a published threat model documenting attack vectors, impact assessments, and mitigations:

Related docs

• SSH Remote Development
• Agent CLI
• Terminal Insights
• App Settings