AI and cyber intrusion to 2028: reading the NCSC forecast
The UK NCSC assesses how AI may change cyber intrusion by 2028. We separate evidence-based forecasts from hype and turn them into a defence plan.
- AUTHOR
- Karol Rapacz / CEO of Breachroad · OSCP · PNPT
- PUBLISHED
- 8 November 2026
- READING TIME
- 18 min read
- TOPIC
- AI Security
The UK’s National Cyber Security Centre published its “Artificial intelligence and cyber intrusion” assessment on 3 November 2026. It forecasts capability through 2028 using a formal probability scale. Its most useful conclusion is not “AI will replace hackers.” Automation will increase the pace, scale and availability of operations, while critical infrastructure remains exposed through known vulnerabilities in ageing systems.
The NCSC assesses that by 2028 nation states will highly likely use AI routinely in cyber operations. It also considers it highly likely that AI-enabled malware will feature in ransomware operations against UK critical national infrastructure. At the same time, the report identifies a new attack surface created by AI systems themselves.
A forecast is not an incident report
This distinction is essential. The NCSC does not say that every scenario has already occurred, nor does it attribute a specific campaign to each judgement. The document is an intelligence assessment using probabilistic language. “Likely” and “highly likely” do not mean certainty.
An organisation should not buy a tool simply because a presentation uses the phrase “autonomous attack.” It should ask which parts of its current defence model weaken when reconnaissance, lure preparation, code analysis and attack iteration become faster and cheaper.
The three changes that matter most
The first is scaling familiar techniques. AI can help tailor phishing messages, analyse public documentation, classify responses and prioritise targets. A defence based on an attacker “not having enough time” becomes weaker.
The second is a shorter interval from disclosure to exploitation. A model can help understand a patch, identify a related pattern and prepare a test. This does not mean every flaw receives a reliable exploit within minutes, but patching processes should be designed for time pressure rather than a relaxed monthly cycle.
The third is attacks on the AI layer itself. Data, models, pipelines, agents, tools, memory and identities add another chain of trust. An adversary can use AI, but it may also influence the model used by a defender or business application.
Why legacy infrastructure remains central
The NCSC highlights known vulnerabilities in legacy critical systems. AI does not need to invent an exotic technique when the Internet still exposes a device with a public panel and months of missed updates. Automated discovery, banner classification and selection of an existing method can increase attempt volume.
Investment in an AI SOC cannot replace an asset inventory, service ownership, segmentation, backups, MFA, contingency plans and telemetry. A tool that summarises alerts cannot patch a device the organisation does not know it owns. External attack-surface management and a mature vulnerability-management programme remain more strategic than chasing one AI feature.
AI in state operations
Routine use may mean analyst support across many stages rather than a fully independent agent. Models can translate documents, analyse code, generate script variants, structure data and propose hypotheses. A human may still select targets, supervise risky actions and judge the value of results.
That matters for defence because detecting “AI-generated content” does not solve the problem. An email written with model assistance is still phishing; model-edited code still performs concrete operations; a login with a stolen token should still violate identity policy. Detect behaviour and impact instead of trying to guess the author’s tool.
AI-enabled ransomware
The report describes a likely direction, not one malware family. AI may improve environment discovery, data selection, analysis of security controls or code adaptation. It does not follow that classic controls stop working. Segmentation, least privilege, resilient backups, EDR and a tested recovery process still remove capabilities from the intruder.
What changes is the response-time requirement. If an adversary moves faster between stages, an alert waiting until Monday has less value. Measure not only MTTD but the time from signal to actually disabling an account, endpoint or segment. Defensive automation is useful when its conditions are explicit, auditable and reversible.
The new AI attack surface
An AI system is more than a model. It includes training and context data, retrieval, the system prompt, a tool broker, memory, tokens, external models and a user interface. Each component creates a distinct trust boundary. Attacks may target instructions, agent permissions, source integrity or key access.
Our AI agent sandbox architecture guide covers tool isolation, while non-human identities for agents addresses short-lived privilege. A mature design combines both with data control and an action ledger. An input filter cannot stop a validly authenticated agent whose permissions are too broad.
A 2026–2028 defence programme
The first workstream reduces existing risk. Inventory edge and critical systems, define the maximum repair time for an exploited vulnerability, remove unused services and test recovery. Every exception needs an owner, deadline and compensating control.
The second secures AI use. Maintain a register of models and agents, define allowed data, test integrations, restrict tools and monitor costs. A model or system-prompt change should be versioned and assessed like any other material software change.
The third improves operational speed. Correlate identity, endpoint, cloud and application evidence; prepare containment actions; exercise outside office hours. Automate repeatable steps while retaining approval thresholds for actions with significant business impact.
The fourth develops skills. Analysts need to verify model output; ML teams need threat-modelling and incident-response fundamentals. Cybersecurity training should use scenarios relevant to the organisation instead of one generic quiz.
Questions for the board
Board members do not need to assess model architecture, but should receive specific answers:
- Which services are critical, and how long can they operate without AI?
- Do we know every public asset and owner?
- How quickly do we remediate KEV items, including at weekends?
- Which agents can take actions rather than only generate text?
- How do we disable a compromised machine identity?
- Can we recover without paying a ransom?
- When did we last exercise a scenario involving both AI and traditional infrastructure?
These questions create a measurable programme. The number of “AI security” licences does not measure resilience.
What to test in a red-team exercise
Testing should cover both adversarial use of AI and vulnerabilities in the AI application. In the first, measure whether faster research and content variants bypass process—not whether a detector recognises model style. In the second, test prompt injection, retrieval, tools, memory, data control and privilege escalation.
The scope requires boundaries, cost limits and a stop procedure. Good AI red teaming ends with evidence, priority and remediation rather than a collection of dramatic prompts. Retesting should prove the control still works after a model change.
NCSC facts and Breachroad conclusions
NCSC facts: by 2028 nation states will highly likely use AI routinely; AI-enabled malware will highly likely appear in ransomware against UK CNI; known legacy vulnerabilities and the new AI attack surface remain important. Breachroad conclusions: the four-workstream programme, board questions and testing approach are our defensive interpretation. We do not present a forecast as a confirmed incident.
The most important next step
Do not try to predict every future technique. Select one critical business process and map its assets, identities, vulnerabilities, AI systems, telemetry and recovery. Then run an exercise assuming the adversary iterates faster than before.
A security audit can establish those dependencies, while secure AI implementation adds control around new components. The NCSC report is not a reason to panic. It is a reason to shorten response times and remove the weaknesses a faster opponent should not discover first.
Working with probability
An intelligence assessment is not a point prediction. “Highly likely” should influence planning, but it does not mean an event will occur in every sector by a precise date. Record the assumption, uncertainty, leading indicators and review date.
Teams might track shrinking disclosure-to-scan time, growth in social-engineering variants and new agent permissions inside the organisation. When an indicator changes, priorities change. The report then remains a working input rather than a universal argument for every purchase.
Scenarios by organisation type
A critical-infrastructure operator should prioritise legacy systems, supplier access, manual control and recovery. A SaaS company focuses on pipeline, APIs, machine identities and tenant isolation. A professional-services organisation needs resilience against tailored phishing, customer-data protection and office-agent control.
The same trend produces a different backlog in each case. Do not copy the NCSC list without context. Begin with a critical business outcome, identify techniques AI can accelerate, then choose controls that limit the consequence.
Safe defensive automation
Defenders can use AI for triage, translation and enrichment. A model should not independently delete an account, isolate an entire segment or publish attribution. Separate recommendation from execution, and use audited thresholds for reversible actions.
Measure false positives, analyst time, decisions corrected by humans and change in mean time to contain. Faster alert closure is not success if hard cases are missed. Evaluation sets need rare incidents and data outside the training period.
Operating during AI degradation
If the SOC or a business process depends on a model, design a mode without it. Define deterministic rules, analyst context sources and procedure locations. Provider failure, model error or suspected manipulation must not stop defence.
Exercise manual triage and restoration. A backlog needs priority so the oldest item does not automatically outrank the most dangerous. The incident-response plan should cover loss of the organisation’s own AI tools.
Quarterly review
Each quarter, compare new evidence with NCSC assumptions: state activity, ransomware, exploitation time, AI-system surface and internal legacy debt. Retire controls that do not reduce risk and move resources as exposure changes. A good AI-cyber strategy is a maintained hypothesis, not a three-year shopping list.


