AI safety

AI safety is the operating model.

We want AI to be useful inside WordPress operations without becoming a black box that quietly changes production under the guise of help.

The goal is not to let AI act first and explain later. The goal is to make the next safe action obvious enough to trust.

Safety principle
Allowed

What the model is supposed to do

The AI should help people think more clearly and move more safely.

Analyze and summarize

AI can read events, correlate change history, and present a clear likely-cause narrative.

Recommend and rank

It can propose next steps, compare risk, and highlight which actions deserve human review.

Prepare a plan

It can assemble a rollback-ready workflow, but the plan still passes through policy gates.

Explain tradeoffs

It can describe impact, confidence, and uncertainty so the operator has context before approving anything.

Blocked

What the model must not do

The product’s safety story becomes real when it refuses the wrong job.

Direct production mutation

AI does not get to freely change live sites just because it inferred the right answer.

Open-ended autonomy

The system does not become a roaming agent that invents its own tasks or skips controls.

Hidden side effects

Actions are scoped, signed, and auditable so the user can see what is happening.

Confidence without evidence

A recommendation without supporting context is a prompt to slow down, not to move faster.

Controls

Safety controls that should stay visible

The product should make the guardrails obvious in the dashboard and in the documentation.

Policy engine

Every recommendation should be evaluated against risk, confidence, rollback readiness, environment, and role permissions.

Human approval

High-risk actions should pause for explicit approval, not quietly slip through a conversational interface.

Failure modes

Safe failure is part of the design

If the model is uncertain or the controls are incomplete, the right answer is to slow down.

Recommend
Default mode
Not execute
Approval
High risk
Human required
Stop
Unsafe state
No silent fallback

Want to see the guardrails in a real workflow?

Start with the docs or a live scan and inspect the safety boundary from the visitor side.