Every AI-generated command must be treated as untrusted input
ven well-intentioned agents execute unexpected package scripts, malformed shell pipelines, or runaway build processes when diagnosing unfamiliar repositories. Running agent tool calls directly on a multi-tenant API host or a developer laptop exposes filesystem secrets, environment tokens, and internal loopback services.
Security cannot be bolted onto prompt instructions alone — model compliance is probabilistic. Containment must be enforced cleanly at the compute boundary.
Filesystem, loopback network, and credential boundaries
Every Helix mission runs inside an ephemeral Linux container or dedicated virtual desktop with its own root filesystem, process tree, and private loopback network namespace. When an agent starts a preview server on localhost, only signed, expiring preview routes reach your browser.
Workspace secrets are encrypted at rest with AES-256-GCM, injected solely into authorized runtime processes, and automatically redacted before logs or transcripts reach storage.
const sandbox = await helix.sandboxes.create({
profile: "linux-isolated",
secretRefs: ["STRIPE_TEST_KEY", "STAGING_DB_URL"],
});
// Secrets exist strictly inside the container process, never echoed in transcripts
const check = await sandbox.exec("npm test -- --runInBand");Stateful durability with instant workspace rewind
Because each workspace maintains clean snapshots alongside exported deliverables, useful build artifacts survive long after ephemeral compute winds down — and entire runs can be rolled back cleanly without leaving stray files behind.
TRY HELIX IN PRODUCTION
Launch an isolated workspace free
Start free on Console or issue tenant-scoped keys inside Developer Console.

