Skip to main content
$ cat SECURITY.md

Security is the architecture.

Not a checklist. Not a compliance page. A 40-page whitepaper, a security.txt, a PGP key, a disclosure policy and a subprocessor list. We would like you to read them.

01

Multi-tenant isolation

Tenant scope is a primary key, not a convention. Every read is bound to the current tenant before it reaches storage. An ACL engine with 334 granular policies, backed by 9 roles, arbitrates every mutation. Cross-tenant leaks are prevented at the query layer — making them architecturally impossible, not just procedurally unlikely.

policies
334
roles
9
leak tests
green
02

Immutable audit

Every write enters a pipeline that produces an audit event, or the request fails. Events are append-only and hash-chained — you can prove a posteriori that no event was altered or removed. Auditors have verified this invariant. The audit cannot be turned off for performance.

events
append-only
hash
chained
invariant
enforced
03

Encryption in depth

Vault uses AES-256-GCM with per-tenant keys. Remote sessions use DTLS-SRTP (the WebRTC-native encryption), with keys rotated per session. TLS 1.3 on the wire. Agents hold per-device keys, revocable by quarantine without reaching the endpoint.

vault
AES-256-GCM
remote
DTLS-SRTP
tls
1.3
04

Identity & access

MFA by email or TOTP. SSO via SAML 2.0 or OIDC with SCIM provisioning. IP allowlists and step-up authentication on sensitive operations. 9 granular roles: 4 MSP-side (MSP_ADMIN, TECHNICIAN, FINANCE, READ_ONLY), 5 client-side (CLIENT_ADMIN, CLIENT_APPROVER, CLIENT_USER, CLIENT_READ_ONLY, CLIENT_GUEST).

sso
SAML / OIDC
scim
yes
mfa
email / totp
05

Agent trust model

Agents enroll via a one-time token, then hold a per-device key pair. No persistent credential lives on the endpoint. Revocation is one call server-side and the agent loses access instantly. Compromised or stolen hosts can be quarantined without requiring endpoint cooperation.

enrollment
one-time
key
per-device
revoke
server-side
06

Data sovereignty

EU-hosted; most AI runs on self-hosted models in our EU datacenter, with a cloud fallback and two features primary on a US provider — every provider is listed on the subprocessors page. GDPR by architecture — subject rights flows built into the ACL engine. NIS2-ready: incident detection, reporting, and recovery workflows documented and tested.

region
EU-hosted
nis2
ready
07

Compliance

ISO 27001-aligned (certification target 2026). SOC 2 Type 2 in progress. GDPR fully aligned. NIS2-ready. DPA provided on request, aligned with EU model clauses. Independent security audits on critical paths (agent, audit pipeline, ACL engine) annually.

iso27001
aligned
soc2
in progress
gdpr
aligned
nis2
ready
08

Responsible disclosure

We publish a security.txt, a PGP key, and a disclosure policy. Researchers receive first-response within 24h and triage SLA within 7 days. We credit responsible disclosure and operate a bug bounty program for researchers we have verified. When we are wrong, we disclose publicly.

security.txt
/.well-known/security.txt
sla triage
24h

Read the full whitepaper.

40 pages. Threat model, data flows, key rotation, disclosure policy, subprocessor list.