What we defend, what we assume, and what we explicitly do not promise. Every control named here exists in the tree; residual risks are labeled as such.
- User source code and project files.
- API keys and tokens (Meta provider key, Tavily key).
- Local app state (settings, sessions, plans, checkpoints, indexes, grants, audit log).
- Integrity of the running app (no silent code execution, no silent updates, no silent provider switching).
┌─────────────┐ preload bridge ┌──────────────────┐
│ Renderer │ ── 51+ typed ────▶ │ Main process │
│ (untrusted │ methods only │ (privileged) │
│ input) │ ◀── JSON data ─── │ services │
└─────────────┘ └──────┬───────────┘
│ fork/utility child
▼
┌──────────────────┐
│ Extension worker │ one per active
│ vm sandbox │ extension; parent
│ inside child │ enforces timeouts
└──────────────────┘ + memory cap
- Renderer → main: every channel checks
isTrustedSenderfirst, then validates payloads withassert*validators. The renderer holds no Node.js, no filesystem, no child processes (contextIsolation,sandbox,nodeIntegration: false). - Extension code → host: extension JS sees only the
kryommembrane (declared commands + permission-gated calls). The membrane runs innode:vminside a dedicated child process; the parent kills the child on timeout, memory-cap breach, crash, or protocol violation. - Child processes: terminal/git/extension children get scrubbed environments (secret keys removed except where the provider call legitimately needs them in the extension worker).
- Network: AI providers, research sources, git remotes, and the update feed. Update URLs must be https and can only come from baked config or operator env — never renderer input.
| Threat | Control | Residual |
|---|---|---|
| Malicious renderer input reaches fs/shell | Sender + schema validation on every channel; workspace containment; path traversal rejection; approval gates for writes/exec | None known; fuzz coverage is manual |
| Path traversal / symlink escape | validate-path containment on every file op |
Symlink races are best-effort (TOCTOU documented) |
| Secret exfiltration via logs/diagnostics/AI | Central allowlist; redacting logger; presence-only booleans over IPC; redacted support bundle | A key pasted INTO a prompt is sent to that provider by design |
| Silent provider/model switching | Explicit selection + explicit fallback only; usedFallback reported |
None |
| Extension steals data / crashes app | Disabled-by-default install; explicit grants; per-call permission checks; isolated child + timeouts + memory cap + crash recovery | NOT a hardened hostile-code sandbox (see below) |
| Malicious update replaces install | No auto-install; https-only feed; strict manifest schema; sha512 verification; staging quarantined; failures clear staging | No code signing yet — owner certs required before public release |
| Supply chain (deps) | Lockfile + npm ci; npm audit in CI; minimal dep set; asar excludes tests/secrets |
No lockfile pin audit service; review is manual |
| Stale AI context causes bad edits | Fingerprinted index; freshness checks; checkpoints + rollback; per-task approval | User can still approve against stale context — shown, not hidden |
- Arbitrary hostile extensions. The child-process + vm + permission architecture raises the bar substantially over in-process vm alone, but it is not an OS-hardened sandbox (no seccomp/AppContainer/job-object confinement, soft memory sampling instead of hard cgroup limits). Do not install extensions you do not trust. Extension installation stays disabled-by-default with explicit user trust for this reason.
- A compromised operating system or Electron/Chromium itself. We track Electron security releases but do not sandbox the main process from the OS.
- AI provider behavior. What Meta (or any configured provider) does with an approved prompt is governed by their policy, not ours.
See SECURITY.md. Private disclosure via GitHub Security Advisories; no public issues for suspected vulnerabilities.