Artificial intelligence is changing how attackers conduct reconnaissance, develop social engineering campaigns, process stolen information and adapt their activity. The most immediate risk for businesses is not necessarily a fully autonomous attack. It is the ability to make familiar attacks faster, more convincing and easier to operate at scale.
The UK National Cyber Security Centre expects AI to increase the frequency and impact of cyber attacks through 2027. It also expects human operators to remain involved in important decisions, even as AI automates individual stages of an attack.
For UK organisations, the practical response is to improve visibility, strengthen identity controls, protect sensitive data and reduce the time between detection and action. Technology can support these outcomes, but it must operate within a clear security operating model.
AI is acting as a force multiplier for attackers. Commercial and open-source tools can help less-skilled operators create more convincing messages, research targets, identify vulnerabilities and process information obtained during an intrusion.
The NCSC’s assessment suggests that the near-term change will be an acceleration of existing attack techniques rather than a complete replacement of human-led operations. AI-assisted reconnaissance, vulnerability research, social engineering and malware development can reduce the time and effort required to move between stages of an attack.
This creates a practical challenge for defenders. Security teams may have less time to investigate an alert before an attacker moves to another stage of the intrusion. Organisations therefore need to combine reliable telemetry, strong identity controls, automated workflows and skilled human oversight.
AI does not need to operate independently to create additional risk. An attacker can use AI to automate selected tasks while retaining control over the decisions that require judgement.
Examples include:
The result is a form of human-machine teaming. Attackers can move more quickly while using human operators to approve important actions or respond to unexpected conditions. Defenders should adopt a similar approach by using automation to reduce repetitive work while retaining human approval for high-impact actions.
AI can reduce the time required for reconnaissance, target selection and exploitation. In some cases, attackers may identify and act on an opportunity before a security team has completed its initial investigation.
This does not mean that every attack happens at machine speed or that human review is no longer useful. It does mean that manual triage cannot be the only line of defence. Security operations need reliable data, prioritised incidents and response workflows that can take proportionate action quickly.
Managed Extended Detection and Response can support this model where an organisation needs continuous monitoring, specialist expertise and structured incident response. The value of MXDR is not simply the presence of a monitoring service. It is the combination of technology, people and processes that helps the organisation identify, investigate and contain threats consistently.
Generative AI can remove many of the language errors that previously made fraudulent messages easier to identify. Attackers can also use publicly available information to create more relevant messages for specific employees, suppliers or executives.
Synthetic voice and video can add credibility to impersonation attempts, particularly when the request involves payment, access, sensitive information or an urgent change to normal process. However, no single technology can determine whether every communication is legitimate.
Organisations should combine:
Microsoft Entra can contribute to this approach through identity protection, conditional access and risk-based controls. These capabilities should support, rather than replace, sound business processes and verification procedures.
Large language models can help attackers produce personalised phishing campaigns at scale. The messages may appear more credible because they use appropriate tone, terminology and context.
Organisations should focus on the consequences of a successful interaction rather than relying only on the quality of the message. Controls should limit what a compromised account can access and make unusual activity visible as early as possible.
Useful measures include:
Data security also matters after an account is compromised. Monitoring unusual data movement can help identify the consequences of a successful phishing attack and support a faster response.
Microsoft Sentinel can bring together security data from Microsoft and third-party sources, helping security teams investigate activity across identities, endpoints, applications, cloud services and infrastructure.
Automation rules and Logic Apps playbooks can support actions such as incident enrichment, analyst notification, ticket creation and selected containment or remediation steps. The appropriate response depends on the incident, the organisation’s configuration and the potential impact of the action.
A practical Sentinel operating model should define:
Sentinel is most effective when it is supported by appropriate data coverage and an operating model that turns alerts into decisions. Deploying the platform alone does not create resilience.
Microsoft Defender XDR can help security teams correlate signals across endpoints, identities, email and cloud applications. This can provide more context than investigating each alert in isolation.
Proactive threat hunting can complement automated detection by searching for suspicious activity that has not yet generated a high-confidence alert. This may include unusual privilege use, persistence mechanisms, suspicious sign-in patterns or activity associated with known attack techniques.
The objective is not to remove the need for alerts. It is to improve the organisation’s ability to identify weak signals, investigate risk and act before an incident causes wider disruption.
Cloud Security Posture Management can also help identify configuration weaknesses in cloud environments. Where AI services or models are connected to business systems, organisations should pay particular attention to permissions, exposed interfaces, data flows and the ability of an AI application to invoke tools or make changes.
AI systems depend on data, which makes data governance an important part of security planning. Organisations need to understand what information is being used, who can access it and how it can be shared with AI applications.
Microsoft Purview can support information protection, data classification, governance, audit and related compliance activities. Its value depends on the quality of the organisation’s data policies, classification approach and enforcement controls.
A practical data governance programme should address:
Model output should not automatically be treated as accurate, safe or authorised. Where an AI system can trigger an action, access data or make a business recommendation, appropriate validation and approval controls should remain in place.
A Cyber Maturity Assessment can help an organisation understand how its people, processes and technology work together. It should identify gaps in areas such as identity, vulnerability management, security monitoring, incident response, supplier risk and data governance.
The proposed Cyber Security and Resilience Bill is intended to strengthen requirements for organisations within scope, including certain essential services, digital services and managed service providers. It is not a general cyber security law applying equally to every UK business, and further detail is expected as the legislation develops.
Organisations should assess their likely obligations rather than assume that the legislation applies to them. They should also treat regulatory readiness as one part of a wider resilience programme. Compliance can provide useful structure, but it does not by itself demonstrate that an organisation can prevent, contain and recover from an attack.
Security leaders preparing for AI-enabled threats should consider the following actions:
AI is changing the economics and operating model of cyber attacks. Organisations should expect faster reconnaissance, more convincing social engineering and greater pressure on security teams. They should also avoid assuming that every future attack will be fully autonomous or impossible to detect.
A resilient security model combines accurate evidence, strong identity controls, protected data, effective monitoring and well-rehearsed response. Microsoft Security can provide an important technology foundation, while specialist expertise can help organisations configure, operate and improve that foundation over time.
CyberOne helps organisations assess their cyber maturity, improve Microsoft Security capabilities and develop practical paths from risk to resilience. The appropriate starting point may be a review of security visibility, identity controls, response processes or AI-related exposure.