The NCSC research problem book
The most significant problems which the NCSC believes are in need of concerted and significant collaborative effort from multiple and diverse disciplines over the next decade.
Page 23 of 29
AI2 - How can AI agents reliably help defenders identify, prioritise and reduce their attack surface?
Sorbetto via Getty Images
AI systems could help defenders understand where they are exposed, which weaknesses matter most, and what actions would most improve their security. This is particularly important as organisations often need to manage large, complex and changing technology estates, where manual approaches to asset discovery, vulnerability management and remediation can struggle to keep pace.
AI agents may be able to support this work by combining information from different sources, identifying likely exposure, predicting which vulnerabilities are most relevant to an organisation, and recommending practical steps to reduce risk. However, reducing attack surface is not simply a technical ranking exercise; it depends on context, operational constraints, business priorities and the consequences of making changes to live systems.
This problem asks how AI agents can help defenders identify, prioritise and reduce attack surface in ways that are accurate, explainable and useful in practice.
Strands or sub-problems
Understanding vulnerability exposure
Organisations may not have a complete or up-to-date view of the systems, services and software they are responsible for. AI agents could help combine information from asset inventories, vulnerability scans, software metadata and other sources to identify where an organisation may be exposed. Research is needed to understand how AI can improve exposure assessment while managing incomplete data, false positives and uncertainty.
Measuring cyber posture
Defenders need ways to understand whether their security posture is improving or deteriorating over time. AI could help develop metrics that combine information about vulnerabilities, exposed services, configuration, asset importance and remediation progress. These metrics need to be transparent and meaningful, so that they support better decisions rather than creating false confidence or encouraging organisations to optimise for the wrong things.
Prioritising risk and vulnerability impact
When a new vulnerability is disclosed, organisations need to know quickly whether they are affected and how urgently they should respond. AI could help link information about vulnerabilities, affected technologies, assets, configurations & dependencies and exposure to estimate likely risk, and estimated impact. More broadly, AI could support prioritisation by helping defenders decide which mitigations are likely to reduce the most risk with the resources available.
Supporting safe system hardening
Many cyber risks can be reduced by changing configuration, disabling unnecessary services, applying mitigations or improving deployment practices. AI agents may be able to recommend,support, or even automate these hardening actions, but changes to live systems can have operational consequences. Research is needed into how AI can suggest practical hardening steps while accounting for system context, operational impact and the need for human approval.
Supporting risk management decisions
Reducing attack surface is predominantly about managing cyber risk, not fixing the largest number of issues. AI agents could help defenders compare options, understand trade-offs and track whether mitigations are reducing risk over time. Research is needed on how AI-generated recommendations should be explained and presented so that human decision-makers can make informed, accountable choices.
Why this is important
With the threat from AI-enabled attackers, which can discover and exploit vulnerabilities faster than ever, it is critical for all organisations to actively manage their attack surface and reduce vulnerabilities as fast as possible. AI can help with many aspects of this, from helping to understand risk, to actively remediating it.
As attack surface management becomes more automated by agentic AI, it is crucial to understand what capabilities are considered mature and which are still early in the research lifecycle, such as autonomous remediation (of not just code but also system configurations). Extensive research is needed before we can trust agents to reliably remediate large parts of the attack surface.
