Preparing for AI-Driven Attacks by 2028: A Practical Security Checklist for UK Legacy Infrastructure

This ItInsightsWorld Security guide converts the latest NCSC warning into an operational deadline for UK critical-infrastructure owners, not a broad forecast about artificial intelligence. Organisations still dependent on legacy platforms need to reduce exposure, patch known defects and improve recovery capability before attackers make routine use of AI-enabled tooling around 2028.

The United Kingdom's National Cyber Security Centre (NCSC), part of Government Communications Headquarters (GCHQ), has recently told operators that hostile use of AI is considered highly likely. The agency's view, also reflected in NCSC guidance, is that AI can lower attack costs and raise attack speed. In recent years, large language and code-analysis models from vendors such as Google, Anthropic and OpenAI have become capable enough to assist with bug discovery and code defect analysis, which originally supports development and defence but can also be exploited by criminals.

Why the warning is a risk-management task

NCSC chief executive Richard Horne argues that new technology is amplifying historical weaknesses. Many organisations spent decades building digital systems at the lowest possible cost, treating security as optional. He warns that if companies and public institutions do not now strengthen basic infrastructure and fix core weaknesses, they can be easily exploited by future attackers.

Additional reporting and NCSC analysis sharpen the target set. The Guardian and NCSC material describe a near-term scenario in which AI-powered malicious software is considered highly likely to be used against UK power stations, airports and hospitals by 2028, while attackers may apply AI capability against known vulnerabilities in critical national infrastructure. For security teams, the practical message is not that AI invents a wholly new attack category, but that it accelerates the discovery and exploitation of old failures.

Threat scenarios and defensive moves

Attack scenario	How AI changes it	Defensive move
Known-vulnerability exploitation	AI may help identify weaknesses and convert code-review findings into exploit paths more quickly.	Inventory external assets, patch known issues, remove obsolete internet connections and enable safe automatic updates where change risk is acceptable.
Large-scale ransomware	AI can assist target research, malware development and faster post-compromise actions.	Segment networks, test offline backups, restore critical services and rehearse ransom scenarios with management.
Phishing and social engineering	AI-generated lure text can become more tailored and available at scale.	Use phishing-resistant controls, verify payments and access changes out of band, train finance, OT and service-desk staff.
Legacy and misconfiguration exposure	AI speeds scanning and analysis, increasing the value of poor hygiene.	Public interfaces should be reduced, not merely monitored; legacy services should be virtually patched or isolated when immediate replacement is impossible.
Legitimate account misuse and remote-service defects	Attackers can use AI to discover weak account paths and remote-service defects.	Review privileged access, disable unused accounts, enforce MFA and monitor anomalous remote access.

A 30-day practical checklist for security teams

1. Map the external surface. List every internet-facing application, VPN, remote-access tool, OT gateway, DNS record and API endpoint. Mark systems that do not need public reachability and close them.
2. Rank legacy systems by consequence. Score each system on operational impact, data exposure, recovery time and replacement feasibility. The highest priority is not the oldest machine by itself, but the oldest machine that also sits on the attack path.
3. Close the known-vulnerability gap. Pull vulnerability reports against the asset inventory, then patch in risk order. Where patching is impossible, apply compensating controls: isolation, virtual patching, reduced privileges and enhanced logging.
4. Reduce phishing and credential pathways. Test whether staff can report and stop a high-quality fraudulent request. Verify payment changes, remote-access grants and service-desk identity checks through an independent channel.
5. Check remote-service quality. Review VPN, RDP, SSH, web-management panels and vendor connections. Many successful intrusions start from ordinary defects, not advanced AI techniques.
6. Run a recovery test. Restore one business-critical workload from backup, validate integrity and record how long restoration takes. If the test cannot complete within an acceptable window, the exposure is not yet under control.
7. Make it a leadership item. Security operations can execute the checklist, but management must set risk appetite, approve funding and hold service owners to deadlines.

Common mistakes that make AI warnings less useful

• Treating AI risk as only a future problem while leaving known vulnerabilities, misconfigurations and exposed legacy systems unchanged.
• Buying more detection tools before removing unnecessary internet connections.
• Assuming legacy assets are too important to touch, and therefore never isolating or replacing them.
• Measuring security only through compliance checklists, instead of testing whether critical services can be recovered after compromise.

UK national response and budget signals

The UK Government has also published funding plans to strengthen national cyber defence. These include a £60 million first phase for the National Cyber Shield programme, referenced as Cyber Shield in the NCSC warning, and £70 million for a National Exposure and Real-Time Vulnerability (NERV) Centre to identify cyber risk across critical infrastructure and the wider economy. A new AI Cyber Security Centre of Excellence in Manchester, operated by the NCSC, is intended to build government AI security capability. The measures are described in UK Government material about the national AI and cyber-security plan.

FAQ: what the warning means for day-to-day security work

Is the threat only about AI-native attacks?

No. The NCSC stresses that attackers often succeed through familiar weaknesses: unpatched known vulnerabilities, configuration errors, legacy systems exposed on the internet, public network interfaces, phishing emails, abuse of legitimate accounts and defective remote services.

Should every organisation stop using legacy systems immediately?

Not always. The safer practical answer is to identify business-critical legacy assets, isolate them, add compensating controls and schedule replacement or segmentation. Where an asset cannot be replaced now, it should not remain unnecessarily internet-exposed.

Who owns the response?

Security teams own execution, but leadership owns risk. Richard Horne's argument is that safety and resilience should be treated as basic operating requirements, not as post-incident repair work. Boards and senior managers must approve the investment required to fix weak foundations.

What should be done in the next 72 hours?

First, publish a list of internet-facing legacy services and confirm whether each one still needs to be reachable. Second, test one offline restore path. Third, brief management that the around 2028 AI-threat window shortens preparation time, because AI can make known flaws and poor hygiene easier to exploit.

Final recommendation

AI is both a defensive amplifier and an offensive enabler. The most credible response is to treat the warning as a deadline for cyber hygiene: reduce external exposure, fix the known gaps, improve recovery, and move security from an optional cost into core operational design. Organisations that only discuss 2028 while leaving today's vulnerabilities visible may find that the future arrives faster than their patch cycle.