AgenturaSecurity & Trust
Security & Trust overview
Agentura runs every customer's agents with hard, kernel-enforced isolation, stores data in the EU, encrypts every credential at rest under a per-company key, and never trains on your conversations. This page summarises the architecture and our security posture.
Per-company isolation (kernel-enforced)
Every company's agents run under a dedicated OS user, on a private container network, with all Linux capabilities dropped and privilege escalation blocked. A deterministic per-company UID plus 0700 permissions mean one company's processes cannot even read another's files. No hypervisor games — plain Unix DAC + network segmentation enforced by the kernel.
Data residency in the EU
Servers and stored data are in Europe; the bundled models run here too. If you choose a third-party model, its prompts go to that provider under their terms — clearly indicated in the product. GDPR-aligned.
Encrypted credentials vault
Connector secrets (API keys, OAuth tokens, passwords) are envelope-encrypted at rest with AES-256-GCM and only decrypted server-side at the moment of an outbound call. Agents never see raw secrets — the platform injects them behind a proxy.
Per-company encryption at rest
Since July 2026, every credential, provider API key, connector token and agent access token is sealed under a data-encryption key unique to your company, wrapped by a platform master key that never touches data directly. Ciphertext is cryptographically bound to your tenant, so one company's records cannot be replayed into another's. Decryption fails closed — a wrong key yields nothing, never someone else's data. Deleting your company destroys its key: cryptographic erasure of every secret it ever stored.
Auditable logging — counts, not content
Every outbound connector call and every model request is logged for audit and metering (provider, model, token counts, cost). The platform records usage counts — not the content of your conversations — and never trains on them.
Hardened egress
Outbound requests are SSRF-guarded: hostnames are resolved and validated against private/loopback/reserved ranges, the connection is pinned to the validated public IP (defeating DNS-rebind), and redirects are not followed. Internal services are unreachable from tenant tools.
Access control
Access is authenticated with short-lived JWTs and scoped to your company. Per-project roles (owner / admin / member) let you compartmentalise access across teams within your organisation.
Independent review
The platform underwent a full security audit in June 2026; all critical, high and medium findings were remediated and verified. Engineering hardening is continuous, on Sysiphus, the agent engine we built and maintain ourselves.