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 22 of 29
AI1 - How can organisations re-design cyber defence workflows to maximise the benefits of agentic AI?
Arslan Haider via Getty Images
Agentic AI systems can plan, use tools and take actions across multiple steps. In cyber defence, they could autonomously carry out tasks such as alert triage, evidence gathering, vulnerability prioritisation, system hardening, detection engineering and incident response.
However, cyber defence workflows are usually designed around human analysts, existing tooling and manual escalation routes. To fully benefit from agentic AI, organisations typically need to redesign these workflows rather than simply add AI agents into current processes.
This problem asks where agentic AI can add genuine value in cyber defence, how it should interact with existing tools, and how humans can retain appropriate oversight and control.
Strands or sub-problems
Designing agentic AI into defensive workflows
Agentic AI will often be introduced into cyber defence processes that were designed for human analysts, established tools and manual decision points. It is important to understand how these workflows should change when an AI system can plan, gather information and suggest or take actions across multiple steps. Research could help identify where agentic AI is most useful, which tasks should remain human-led, and how organisations can design workflows that significantly improve defensive outcomes.
Safe use of existing cyber defence tooling
Many cyber defence teams already rely on monitoring platforms, vulnerability scanners, ticketing systems and configuration tools. Agentic AI systems may need to interact with these tools to be useful, but doing so can create new risks if access is too broad or actions are not properly controlled. Research is needed to understand how permissions, approval points, logging and recovery should work when AI agents are connected to operational cyber security tooling.
Human oversight and shared control
Cyber defence often involves decisions that depend on context, uncertainty and operational judgement. Agentic AI may be able to support these decisions, but it should not remove the need for meaningful human oversight where it matters. Research could help define when human approval is essential, how operators should review or challenge AI-recommended actions, and how people can maintain situational awareness without creating unnecessary bottlenecks or approval fatigue.
Single-agent and multi-agent approaches
Some cyber defence tasks may be suited to a single AI agent, while others may benefit from multiple specialist agents working together. Multi-agent systems could improve coordination and scale, but they may also make behaviour harder to understand, monitor and control. Research is needed to understand when multi-agent approaches add value, how agents should communicate, and how organisations can manage disagreement, error propagation and the risk of compromised or misleading agent-to-agent messages.
Why this is important
Safe and effective adoption of agentic AI into cyber defence workflows is a challenge facing organisations today. Organisations are exploring the use of AI to increase the security of their systems, the detection of adversaries and the speed of their response and recovery, while balancing the risk of those agents introducing security issues themselves, or causing unintentional harm to themselves or others. Solving this challenge is fundamental to the future of cyber security. Longer term, to realise the full benefits, we will need to redefine our cyber workflows to be agentic-centric, allowing our agents and systems to work together in support of our defence whilst retaining meaningful human oversight, accountability and trust. This is the ultimate ambition of our Cyber Shield initiative and research will be essential to getting us there.
