The National Cyber Security Centre published its 2026 Annual Review on 3 November, covering the twelve months from 1 September 2025 to 31 August 2026. Most of the coverage led with artificial intelligence, and the AI warning is real. But the most useful line in the review is quieter, and it is about doors: for large organisations, 88% of incidents came through just four initial access routes.
That is good news, of a kind. Four doors are a manageable number to close.
The numbers, and how to read them
Over the year, the NCSC's incident management team received 1,831 reports of cyber attacks. Of those, 337 needed the NCSC's direct support, and 80 were rated nationally significant. The review also cites government analysis putting the cost of malicious cyber activity to UK organisations at an estimated £14.7 billion a year.
The tempting headline is that nationally significant incidents fell sharply from the previous review. Treat that with care. The review doesn't explain the drop or say whether the way incidents are categorised has changed, and serious incidents have always arrived in clusters: one bad quarter, or one widely exploited vulnerability, moves the total a long way. A quieter year for the NCSC's most serious category says little about your organisation's odds.
The four doors
The review names the four routes behind 88% of incidents at large organisations:
- Exploiting public-facing applications: a vulnerable web app, portal or appliance on the internet.
- Phishing: a message that tricks someone into handing over access or running something.
- Misuse of valid accounts: an attacker signing in with real credentials, stolen or bought.
- Exploiting external remote services: VPNs, remote desktop and other ways in from outside.
None of these is new, and that is the review's central point. The NCSC's argument is that AI isn't inventing new methods of attack. It is letting attackers find and exploit weaknesses that organisations have tolerated for years, faster and at a scale that wasn't previously practical. The review notes that the bulk of recent state-backed activity is "a much simpler exploitation of common vulnerabilities and exposures", not an advanced, autonomous AI attack.
The AI warning
The forward look is where the review is most pointed. The NCSC judges it highly likely that, between now and 2028, AI-powered malware will have been used in a ransomware attack on UK critical national infrastructure. It also expects China, Russia and other states to be routinely using AI in cyber operations by 2028.
If you run critical infrastructure or supply it, that is a direct planning assumption for the next two budget cycles. If you don't, the same logic still applies one step removed: tooling built for high-value targets gets cheaper and spreads. The defence is the same either way, because AI-assisted attackers still have to come through one of the four doors.
Closing the doors, in order
Each door has a well-understood control, and most organisations already license the tools for all four.
1. Public-facing applications: know what you expose, and patch it first. You can't defend an asset you don't know is on the internet. Build and maintain an inventory of every internet-facing system, including the forgotten test portal and the appliance a supplier installed. Then patch those systems on a shorter clock than everything else. In Microsoft estates, Defender External Attack Surface Management and Defender Vulnerability Management do much of the discovery and prioritisation.
2. Phishing: make sign-in phishing-resistant. Training helps, but it doesn't scale against well-written, AI-generated lures. The control that does is sign-in that can't be phished: passkeys, security keys or Windows Hello for Business, rather than passwords plus codes that can be relayed. Pair it with mail filtering and a one-click way for staff to report anything odd.
3. Valid accounts: watch behaviour, not just credentials. When an attacker signs in with a real password, the login itself looks normal. What gives them away is what happens next: an impossible journey, an unusual device, a sudden interest in the finance share. That is the job of identity threat detection and response, and it applies to non-human identities, such as service accounts and AI agents, as much as to people. Conditional Access, Entra ID Protection and privileged identity management are the Microsoft building blocks.
4. External remote services: retire what's exposed, gate what's left. Remote desktop open to the internet should simply not exist in 2026. VPN appliances are a favourite target because they sit on the edge and are slow to patch. Where you can, move remote access behind identity-aware access, such as Entra Private Access or another zero-trust network access product, so every connection is checked against who is asking and from what device.
None of these four is a project in its own right. Together they are the core of a zero-trust programme adopted in steps, and the order above is a sensible order to take them in.
A one-page check for your next board meeting
If you want to turn the review into a short conversation with your board, ask four questions, one per door:
- Do we have a current list of everything we expose to the internet, and how quickly do we patch it?
- What share of our sign-ins are phishing-resistant today?
- Would we notice a valid account behaving unusually, and how fast?
- Is any remote access still open to the internet without identity checks?
If any answer is "we're not sure", that is where to start. Our zero-trust readiness scorecard takes about five minutes and maps your answers to the same priorities.
The bottom line
The NCSC's 2026 review is not a call to buy something new. It says that attackers, with AI behind them, are getting faster at walking through the same four doors they have always used. The organisations that come off worst will be those that left those doors open. If you'd like a second pair of eyes on yours, our cybersecurity team works through exactly this, starting with the Microsoft security you already pay for.

