Safety and Guardrails

Quox's safety system keeps you in control. With kill switches, approval workflows, staged rollouts, and comprehensive audit logging, you get the power of AI automation with complete peace of mind.

Why safety matters

When AI manages real infrastructure, the stakes are high. A misunderstood command could restart production services, a misconfigured script could affect dozens of servers, or a simple typo could cascade into an outage.

Quox's safety system exists so you can confidently delegate to AI knowing that guardrails are in place. You get the speed and intelligence of AI automation while retaining complete control over what actually happens.

You are always in control

Quox's safety architecture is built on one principle: nothing happens without your awareness and consent. Here is how that works in practice:

The kill switch

A single button stops everything. Immediately.

Think of it like an emergency stop button in a factory. You hope you never need it, but knowing it is there lets you work with confidence.

Dry-run mode

Not sure what a command will do? Run it in preview mode first.

In dry-run mode, Quox plans and analyzes but does not execute. You see exactly what would happen without any actual changes to your infrastructure.

When to use dry-run:

• Testing new workflows before going live
• Reviewing AI-generated plans before approval
• Training new team members safely
• Auditing what automation would do

DEFCON safety levels

Quox uses a DEFCON-style system to classify operations by risk. Higher numbers are safer (just like the real DEFCON scale).

Level overview

How levels work

GREEN and BLUE (Automatic): These are read-only or low-risk operations. Quox handles them without interrupting you. Perfect for monitoring, reporting, and routine health checks.

AMBER (You Approve): Before anything runs, you see a plan with full details. One click to approve, one click to reject. The workflow waits as long as needed for your decision.

ORANGE (Two-Person Rule): Critical operations require two approvers. This prevents any single person (or the AI) from making high-impact changes alone.

RED (Manual Override): Some operations are too sensitive for automation. Quox will not even attempt them. You must do these yourself, deliberately and carefully.

Configuring your safety level

You can adjust the default DEFCON level based on your comfort and needs:

// In your Quox settings
{
  "default_defcon": 4,     // Start at BLUE for most operations
  "require_approval_below": 3,  // Require approval at AMBER and below
  "two_person_below": 2,   // Require 2 approvers at ORANGE and below
  "disabled_below": 1      // Disable automation at RED
}

Recommendation for new users: Start at DEFCON 3 (AMBER). This means you approve most operations while still getting AI assistance with planning. As you build confidence, you can shift to DEFCON 4 or 5 for routine tasks.

What gets blocked (and why)

Quox maintains a denylist of commands that are never executed automatically, regardless of DEFCON level. These require explicit manual action.

Always blocked

Blocked without explicit confirmation

Example: what happens when something is blocked

Say you ask Quox to "clean up old logs on all servers." Here is what happens:

1. Quox plans the operation
2. Detects it affects more than 5 hosts
3. Pauses and asks: "This will affect 47 hosts. Please confirm multi-host operation."
4. You review the exact command and host list
5. Only after your confirmation does it proceed

Nothing surprising. Nothing hidden. You see everything before it happens.

Approval workflows

For AMBER and higher-risk operations, QuoxFlow uses a structured approval flow with approval requests delivered directly to your Inbox.

How approvals work

You request an action
       ↓
QuoxFlow evaluates policies (with inline testing)
       ↓
Policy triggers approval requirement
       ↓
Approval request appears in your Inbox
       ↓
You review: commands, targets, expected outcome
       ↓
Approve → Executes with full evidence capture
Reject → Cancelled, nothing happens

What you see in an approval request

Inbox approval features

• Integrated Inbox - Approval requests appear in your QuoxCORE Inbox alongside other notifications
• Policy management - Full CRUD for policies with inline testing before deployment
• Mobile friendly - Approve from your phone via Slack, Telegram, or email
• Time limits - Set expiration for time-sensitive approvals
• Escalation - Auto-escalate to backup approver if you are unavailable
• Evidence trail - Every approval/rejection is logged with timestamp, approver, and linked to execution evidence

Staged rollouts (canary deployments)

For multi-host operations, Quox uses staged rollouts to catch problems early.

How it works

Instead of running on all hosts at once, Quox deploys in waves:

After each wave: Quox checks health. If anything fails, the rollout stops immediately. No cascade failures.

Example scenario

You want to update a package on 50 servers:

1. Wave 1: Updates server-01. Waits. Checks health. All good.
2. Wave 2: Updates server-02, server-03, server-04. Checks health. All good.
3. Wave 3: Updates 10 more servers. Checks health. server-12 fails health check.
4. Stops immediately. 36 servers untouched. You investigate server-12.
5. Your decision: Fix the issue and resume, or roll back the 14 affected servers.

You caught a problem at 28% instead of discovering it after 100%.

Enhanced audit trail

Every action Quox takes is recorded with full traceability. Nothing is hidden or lost. QuoxFlow's enhanced audit trail adds correlation ID search, type filtering, and causality chain tracing across workflows and agent handoffs.

What gets logged

Accessing the audit log

• Evidence Browser: Search by correlation ID or intent, with detail panels and causality chain views
• Execution Drilldown: Click any execution for APM-style waterfall visualization with collapsible payload data
• Running Now Monitor: Live dashboard showing active executions with animated progress bars (5-second polling)
• Schedule Visibility: See all scheduled/cron workflows with next run times, countdown timers, and execution history
• API: Query the full audit log programmatically by correlation ID or type
• Export: Download logs for compliance or analysis

Retention

Audit logs are retained for 90 days by default. Critical events (approvals, blocks, safety activations) are retained indefinitely. Evidence records are immutable and queryable for the full retention period.

Preflight and postflight checks

Quox verifies conditions before and after every operation.

Before running (preflight)

If any check fails, the operation is paused and you are notified. No surprises.

After running (postflight)

If postflight checks fail, Quox can automatically trigger rollback (if configured) or alert you for manual intervention.

Best practices

For new users

1. Start with DEFCON 3 (AMBER) - Approve everything until you trust the system
2. Enable dry-run for new workflows - Preview before committing
3. Review the audit log regularly - Understand what Quox does
4. Set up mobile notifications - Stay informed even when away from your desk

For production environments

1. Use two-person approval for production - No single point of failure
2. Enable staged rollouts - Catch problems early
3. Configure escalation policies - Approvals should not block on one person
4. Set up backup approvers - Coverage during time off

For high-security environments

1. Set DEFCON 1-2 as default - Require approval for nearly everything
2. Use RED level for sensitive systems - Manual only for critical infrastructure
3. Enable comprehensive audit logging - Keep everything for compliance
4. Regular access reviews - Ensure only authorized users can approve

Quick reference

Summary

Quox's safety system is designed to give you confidence, not to limit you. With the kill switch, DEFCON levels, approval workflows, staged rollouts, and comprehensive audit logging, you have complete visibility and control over every action.

The goal is simple: You should feel comfortable letting AI handle your infrastructure because you know exactly what it will do and you can stop it at any time.

Safety is not a limitation. It is what makes powerful automation trustworthy.

"I'm sorry Dave, I'm afraid I can't do that... without proper approval."