Security

Autonomy with adult supervision.

FrontBrain AI is designed around controlled execution: approval gates, Vault-backed secrets, connector scoping, read-before-write, workspace isolation, audit logs, and verification-proof output.

Autonomy is useful. Ungoverned autonomy is how things go sideways.

Platform brakes

The platform enforces the brakes.

Intent-based mutation gate

Risk is classified by action impact. Local artifacts pass freely. Deployments, DNS, firewall, email, billing, production installs, database writes, and power actions can require approval.

Read-before-write

A remote write cannot execute until FrontBrain AI has inspected the same target in the same session.

Stall detection

Repeated same-error loops, no-state-change patterns, and platform-gate loops pause execution before wasting time or credits.

Approval workflows

Approval can be per call, per campaign, per workspace, or multi-party for high-risk operations.

Append-only audit log

Every connector touch, tool invocation, approval, output, and verification result is logged.

Verification-proof completion

Tasks do not claim “done” without evidence.

Secrets & data

Where your data lives, and where it does not.

Vault for credentials

Connector credentials live in Vault and are scoped by user, workspace, and target.

Per-tenant isolation

Conversation history, audit logs, artifacts, and connector access are tenant-scoped.

MSP workspace isolation

Each client workspace can have separate connectors, users, permissions, and audit exports.

BYOK support

Use your own model keys where supported, keeping usage under your terms, billing, and retention policies.

Encryption in transit

TLS everywhere. Remote Gateway uses outbound WebSocket-over-TLS.

Encryption at rest

Platform storage and artifacts use encrypted storage. Private deployments can use customer-controlled storage.

Governance

The controls your security team will ask about.

Role-based access

Tenant admins, users, connector scopes, and tool allow-lists.

SSO

Google SSO for Business+.

Audit exports

Export by tenant, workspace, user, connector, date range, or task.

Connector-scope gating

A tool targeting a connector the user does not own is rejected.

Planning mode

High-risk requests start in read-only planning mode until approved.

Reproducible audit trail

Plans, tool calls, approvals, outputs, and verification evidence can be reviewed after the fact.

Talk to us about security
Ready?

Give autonomy a spine.

Every important action has a reason, a result, and an evidence trail.

Less dumb. More done.