← Back to homeTrust & security
How we protect your data and keep AI actions safe. Every measure below maps to a real control in the platform.
🧱Isolation & network
- 📦
Sandboxed code execution
Untrusted code runs in a gVisor sandbox: no network, read-only filesystem, all capabilities dropped, CPU/memory/time caps.
- 🌐
Controlled egress
All outbound traffic goes through a single audited proxy with an allow-list and SSRF protection.
- 🔒
Locked-down backends
Internal services are loopback-bound and require a shared internal token; the model never sees your secrets.
🔑Identity & access
- 🎫
Strong sessions
RS256-only JWTs with audience/issuer/expiry checks; logout and password reset revoke every existing token.
- 🏢
Enterprise SSO
Sign in with Google or your own identity provider (OIDC — Okta, Azure AD).
- 📝
Audit trail
Sensitive actions — logins, shares, approvals, admin changes — are recorded in an append-only audit log.
🗄️Your data
- 🧩
Tenant isolation
Knowledge bases and documents are scoped per organization; one tenant can never read another's data.
- 📤
Export & deletion
Download all your data at any time, or delete your account — with a scheduled hard purge (GDPR).
- ⏳
Configurable retention
Admins can set how long conversations are kept; older ones are purged automatically.
🤖Safe AI actions
- 🛡️
Prompt-injection defense
Once untrusted content is ingested, side-effect and exfiltration tools are blocked (lethal-trifecta defense).
- ✋
Human-in-the-loop
Actions with real-world effects (email, publishing) can require explicit human approval before they run.
- 📊
Spend controls
An atomic, non-negative credit ledger plus per-organization budget limits keep usage under control.
This page describes controls in place today; it is not a certification. Formal SOC 2 / ISO 27001 evidence is in progress.