Skip to content
  1. Home
  2. Articles
  3. Before a System Fixes Vulnerabilities in Production by Itself: Approval Tiers, Scope and a Stop Switch to Agree First
114
Security

Before a System Fixes Vulnerabilities in Production by Itself: Approval Tiers, Scope and a Stop Switch to Agree First

Before a System Fixes Vulnerabilities in Production by Itself: Approval Tiers, Scope and a Stop Switch to Agree First

Gartner lists autonomous exposure remediation as a 2027 trend. Before a tool may patch production on its own, agree approval tiers, scope, a stop switch and the contract wording.

A queue that never empties

Your security team opens the monthly scan report and finds several hundred open rows. Most belong to other teams and wait for a maintenance window that may be three weeks away. This is not a skills problem. It is a volume problem: new findings arrive faster than people can read them, decide, test, and negotiate downtime.

That is why letting a tool fix things on its own is tempting. If a system can read a finding, rank it, test a patch and apply it the same night, the queue shrinks. But automatic remediation adds one new thing: a tool with write access to production and no human pressing confirm each time.

The queue is never empty.

This article does not ask whether the technology is good. It asks what an organisation must agree before it hands that access to a tool.

What Gartner says, and what the page does not prove

Network World, in an article by Ann Bednarz dated 20 October 2026, reports that Gartner lists autonomous exposure remediation (AER) among its 2027 trends. Gartner defines it as an integrated set of capabilities that combine AI techniques and automation to find, prioritize and fix security vulnerabilities across the enterprise digital estate. The article says such a system automatically tests and deploys patches or virtual patches into production. We read the article ourselves on 12 November 2026. Gartner's own page was not available to us, so we cite only the Network World report.

Two numbers need care. Gartner is quoted as saying the traditional defensive process is broken, leaving organizations with an average 50-day exposure window. The page does not say where the 50 days comes from, so we report it as Gartner's statement, not as our own finding.

The second is a forecast. Gartner predicts that by 2031, organizations deploying AER will reduce their average vulnerability containment window from weeks to within 48 hours. That is a prediction for five years out, not a measurement. The page does not define "containment window" or say how many weeks the baseline is. Do not write 48 hours into a service level you can sign today.

Before the tool: an owner list and approval tiers

An inventory where every system has an owner

An automated fixer can only act on what it knows exists. The first precondition is an asset list in which each system has a named person as owner, with contact details, not a department name. Without an owner, nobody can approve or refuse a change, and nobody answers the phone at 2 a.m. when a patch stops a service. If you cannot say who owns a given server, fix that before you evaluate any product.

The record should hold more than a hostname: the services running on it, the systems that call it, the window in which it may be taken down, a backup contact for when the owner is away, and whether a vendor manages it. The tool reads these fields to decide whether a fix is allowed, so a record that does not match reality makes the tool decide from a wrong picture.

Three tiers

We suggest three tiers, written down and signed by system owners. Tier one: fixes may run unattended. Tier two: a named human approver must say yes each time. Tier three: never automated, handled by a person through the normal change process.

Four criteria place a system in a tier. Blast radius: how many users and systems suffer if the fix goes wrong. Reversibility: how quickly it can be undone, and whether data written in the meantime survives the undo. Dependency: whether other systems rely on it, since a patch that changes one library's behaviour can ripple down the chain. Control: whether a vendor or a contract decides how changes are made. If a vendor controls change, the organisation has no right to let a tool touch it.

As a starting point for debate, a rebuildable internal test host could sit in tier one. A public web tier behind a load balancer might sit in tier two so that a person watches the canary. A core financial database, anything tied to control equipment, and anything a vendor holds change rights over belong in tier three. Placement is the system owner's decision; these are examples, not rules.

What a virtual patch does and does not do

A virtual patch is a control placed in front of a flaw to block attacks against it, such as a web application firewall rule or a block rule on a network device, without changing the vulnerable software. It is quick and usually easy to remove. It makes sense when the real patch has not been released or the system cannot be taken down yet.

What it does not do is remove the flaw. The vulnerable code is exactly where it was. The control only lowers the chance that outsiders reach it. An attacker who finds another route to the same spot, from inside the network or through a stolen account, may bypass it, and a rule written too narrowly leaves gaps.

Every virtual patch needs an expiry date. Record when it was applied, which flaw it covers, and the date by which a real fix must replace it. Without that, a temporary control stays forever, and the asset list shows the flaw as handled while it is not fixed.

We therefore suggest two separate counters on the report: flaws reduced by a temporary control, and flaws fixed at the source. An executive who reads one blended number may believe more is fixed than is.

Proof before production, and a stop switch

Evidence first

Automation does not waive release discipline. Before touching production, the tool should show three things: test results from an environment close to the real one, a staged plan such as canary or phased rollout, and a rollback plan that has been tried. Deployment strategy and change approval are covered in our article on production change and release management. The point to add here is that your freeze windows must bind the tool exactly as they bind people.

One switch that halts everything

There must be a single switch that stops all automated fixes, and four things must be agreed beforehand. Who may press it: we think at least two roles reachable around the clock, for example the person accountable for security and the person accountable for operations. How fast it takes effect. What happens to work in flight. And who decides when to turn it back on.

The point most often missed is the half-applied change. If the tool has patched 40 of 100 machines when it is stopped, you are running two versions side by side. The stop procedure must say whether to roll back, leave as is, or hand the decision to a person. Rehearse it outside production, and place it in your outage roles and decisions plan, because a misbehaving tool is one of the ways an outage starts.

A trail auditors can read

Every action needs a record that an auditor can read without relying on the tool itself: the finding that triggered it, the tier applied, the approver's name for tier two, test results, start and end times, and the outcome. Keep it where the tool can write but cannot edit or delete.

What stays out of scope, and what to write in the TOR

Legacy and vendor-managed systems

Two groups should be out of scope from the start. Legacy systems with no clear owner, because nobody can approve and nobody knows what a patch will collide with. And systems a vendor maintains under contract, which usually say changes go through the vendor. Letting a tool alter them may void the vendor's warranty or responsibility. These systems still need other protection, such as network isolation or extra monitoring, just not through the automated fixer.

Contract wording for an operated service

If a provider will run such a tool for you, the TOR should cover four things. First, the list of systems in scope and the tier of each, attached as a controlled document. Second, read access to the tool's change log at all times. Third, a kill switch held on your side that you can press without waiting for the provider. Fourth, liability for damage caused by automated changes, stated plainly: which party answers when a patch the tool applied takes a system down. Your legal counsel should set the exact wording and any amounts. For general TOR conditions that keep a system maintainable, see our article on TOR conditions for a maintainable system.

Limits we should state plainly

If your backlog is a few dozen items a month, a human-run patch cycle is enough. The case for AER is volume and a shortage of skilled people. If neither describes you, automation removes little work and adds a new set of things to run: write access to production, approval tiers, a stop switch and an audit trail. That is a monthly cost, and it is a new thing that can break production.

There is also the evidence. The 50-day and 48-hour figures are Gartner statements as reported by Network World. We have not seen an independent measurement showing any organisation reaching them.

Objections we often hear

A common objection is that attackers already automate, so defenders cannot wait for a person to click approve. We agree speed matters. The proposal here does not put a person on every fix. Tier one is designed to run without waiting. Only systems where a mistake costs a lot wait for a human, and you decide how wide tier one is, widening it as real operation proves the tool does not break things.

Where to start

Start with what needs no product. Take your ten most important systems and ask whether each has a named owner. Then sit with those owners and place each system in a tier using the four criteria. Then write down who can press the stop switch and what happens to work in flight. If you cannot answer these, evaluating tools is premature, since even a good tool would have no frame to work inside. For the scanning and testing that should run alongside, see our penetration test procurement guide.


Summary

Before a tool fixes production by itself, be able to say which systems it may fix alone, which need a named approver, which it must never touch, and who can halt everything within how many minutes. If you cannot, the problem is not the tool.

If you would like to talk through the details of a project, you can reach us at 088-983-9386 or [email protected]. Our office is in Bang Kapi, Bangkok. For organizations still drafting a scope of work, we are glad to arrange a conversation to understand the problem first, without needing complete documents. And if we find that the problem is not yet at the point of needing a new system, we will say so plainly, because starting a project at the wrong time costs more than waiting another quarter.

FAQ: Frequently Asked Questions about This Article

A collection of questions and answers to help you better understand the content of this article.

Tags:

autonomous exposure remediationautomated vulnerability remediationvirtual patchingpatch approval processkill switch automationvulnerability management TORGartner top tech trends 2027exposure management
Share:

Other Articles

Stay tuned for upcoming articles!