Safety by Design

What if the AI does something stupid?

The same question every industrial buyer asks. Here's how we answer it.

Air-gapped · EU AI Act-aligned · Five-tier safety authorization · Full audit trail.

5-Tier Authorization

Every AI action has a tier. Every tier has an authorization rule.

Brain's AI agent operates under a strict 5-tier model. Read operations are always allowed. Safety-critical actions require admin approval AND explicit confirmation. No exceptions, no shortcuts, no black boxes.

0Tier

Category

Read

None

Authorization

Always allowed

Example

What's the current temperature?

1Tier

Category

Inform

None

Authorization

Always allowed

Example

Show me the alarm log

2Tier

Category

Configure

Low

Authorization

Agent enabled (default)

Example

Create a new tag for tank 3

3Tier

Category

Control

Medium

Authorization

Operator approval or pre-approved rule

Example

Start the heater at 60°C

4Tier

Category

Safety-Critical

High

Authorization

Admin approval + confirmation

Example

Disable the overpressure alarm

All five Brain agents (Engineering, Commissioning, Watchdog, Responder, Optimizer) operate under the same tier model — see /agents.

Compared to Other AI-PLC Platforms

Brain is the only AI-PLC platform with a published safety architecture.

Siemens TIA Portal Eigen has no published authorization model. Rockwell FactoryTalk Studio Copilot relies on standard RBAC. Beckhoff CoAgent has no equivalent tier system. CODESYS MCP defers all safety decisions to the host application. Brain’s 5-tier model with auto-lock and emergency stop is unique in the market today.

Auto-Lock

Three Tier-3 actions in 60 seconds. Auto-lock engages.

If Brain's AI agent executes 3 or more Tier-3 actions in 60 seconds, all autonomous modes are disabled automatically. Requires manual re-enable by an admin. This prevents runaway loops and accidental cascading changes.

T+0s

Tier-3 action #1

Within 60s window

T+18s

Tier-3 action #2

Within 60s window

T+47s

Tier-3 action #3

Within 60s window

Result

All autonomous modes disabled

Requires manual admin re-enable.

Emergency Stop

One endpoint kills everything.

POST /ai/emergency-stop. Kills all agent activity instantly. Cancels pending actions. Disables all autonomous modes. Notifies all operators. One call, full shutdown.

curl · emergency-stop.sh

POST https://brain.local/ai/emergency-stop
Authorization: Bearer YOUR_ADMIN_TOKEN
Content-Type: application/json

{
  "reason": "Unusual behavior detected",
  "disable_until": "manual"
}

// Response
HTTP 200 OK
{
  "status": "all_autonomous_disabled",
  "actions_cancelled": 7,
  "operators_notified": 4,
  "timestamp": "2026-04-23T14:32:08Z"
}

Audit Trail

Every action. Every approver. Every timestamp.

Every AI action is logged to SQLite with: the tool used, the arguments, the tier, who approved it, when, and what the result was. Queryable via admin dashboard. Exportable for compliance audits. Immutable.

audit_log.sqlite · event #4f82a1c9e

IMMUTABLE
{
  "id": "evt_4f82a1c9e",
  "timestamp": "2026-04-23T14:18:42.193Z",
  "tool": "plc.set_setpoint",
  "tier": 3,
  "args": {
    "tag": "HEATER_01.setpoint",
    "value": 60.0,
    "unit": "celsius"
  },
  "approved_by": "operator:mkowalski",
  "approval_method": "pre_approved_rule:rule_88c",
  "result": "success",
  "duration_ms": 142,
  "previous_value": 55.0,
  "session_id": "sess_9a3b7e2"
}

Degraded Mode

What happens when something fails.

AI service down

PLC keeps running.

Backend crash

PLC keeps running.

Network outage

PLC keeps running, alarms still fire locally.

Cabinet power loss

PLC restores last known state on power-on.

The PLC engine, hardware drivers, and alarm processing all run independently of the AI agent. See /reliability for the full degraded-mode behavior.

Approve & Remember

Day one: the agent asks permission. Day thirty: it runs autonomously.

When an operator approves a Tier-3 action, they can select “Approve & Remember” to create a pre-approved rule. Future identical situations are handled without re-prompting. The agent learns YOUR safety boundaries.

Day 1

Agent asks permission

Pending approval · Tier 3

Start heater HTR_01 at 60°C?

Tank TK_03 temperature low · warmup sequence required

DenyApproveApprove & Remember ✓

Operator clicks Approve & Remember. A pre-approved rule is created.

Day 30

Agent runs autonomously

HTR_01 → 60°C (auto, rule:88c)

HTR_02 → 45°C (auto, rule:88c)

PUMP_03 ON (auto, rule:91a)

VALVE_V7 → 45% (auto, rule:91a)

HTR_01 → 62°C (auto, rule:88c)

MIXER_M2 → 180rpm (auto, rule:94e)

+ 14 more approved operations...

20+ operations executed autonomously. Every one logged, auditable, and revocable.

Limits and Revocation

Rules are revocable, expirable, and auditable.

Every Approve & Remember rule is stored with the operator who approved it, the conditions that match, and an expiration date (default 90 days, configurable). Rules can be revoked instantly by any admin. Every autonomous action under a rule is logged with the rule ID for full traceability. Rules cannot escalate above Tier 3 — Tier 4 always requires fresh admin approval, no remembering allowed.

Change Management

Propose. Review. Approve. Deploy.

PLC changes go through a Git-native workflow. Every change is proposed as a diff, reviewed by authorized engineers, approved with signature, and deployed with one-click rollback. Zero unreviewed PLC changes. Ever.

01

Propose

AI or engineer proposes a PLC change as a Git diff. Every changed line is annotated with intent and impact analysis.

02

Review

Authorized engineers inspect the diff, run it in simulation, and see automated impact analysis: which alarms could fire, which HMI screens reference these tags, which recipes are affected.

03

Approve

Signed approval required before deploy. Cryptographic signatures stored in audit trail. Approval can require multiple signoffs.

04

Deploy

One-click deploy with instant rollback. Previous version is kept available for 90 days. Zero unreviewed changes ever reach the PLC.

Safety Stats

The numbers that matter.

5

Tiers of authorization

2s

Hardware watchdog timeout per I/O board

100ms

PLC scan cycle (continues independent of AI)

0

Unreviewed PLC changes

Immutable audit trail

Incident Response

If the worst happens.

Every Brain cabinet has an external emergency stop button physically wired to a hardware safety relay — independent of the controller, independent of the AI. The hardware E-stop cannot be overridden by software. AI safety failures are escalated to operators within 1 second via local alarm and configurable notification channels (email, SMS, webhook). Post-incident, the full audit trail provides a complete forensic record for root-cause analysis.

Safety Roadmap

What’s shipped, what’s in progress.

Today

5-tier authorization, auto-lock, emergency stop, audit trail, Approve & Remember, change management workflow, hardware watchdogs.

In Progress

IEC 62443 industrial cybersecurity certification (Q4 2026), EU AI Act conformity assessment package (2027), IEC 61508 functional safety qualification (target 2027), CE marking (2026).

(Certification timelines published when available.) See /reliability for the broader reliability roadmap.

Verify, Don’t Trust

How to verify our safety claims.

Safety is not a feature. It's the foundation.