How Boards Should Govern AI and Agentic AI Cyber Risk
The arrival of generative AI was, for most boards, a productivity question. The arrival of agentic AI systems that take actions on an organisation's behalf is a governance question, and a harder one. When a model can read your data, call your tools and execute transactions, the attack surface is no longer a network; it is a set of decisions your organisation never explicitly made.
This article sets out a board-level framework for governing AI and agentic AI risk, focused on the decisions only the board can take.
From tool to agent
The first shift to internalise is behavioural. A language model that drafts text can be governed like any software: approved, monitored, switched off. An agent that books, pays, emails or trades on your behalf behaves. It takes actions in the world, and those actions can cause harm, breach a contract or violate a regulation.
Boards that govern AI only as a tool will miss the risk that actually matters: an autonomous system taking a decision no human reviewed, faster than any human could intervene. Governance has to account for behaviour, not just output.
AI governance as a board responsibility
The EU AI Act, ISO/IEC 42001 and the NIST AI RMF each converge on the same expectation: AI risk is governed from the top, with documented accountability, a defined risk appetite and human oversight. For the board, that means three things:
- A clear statement of where AI may be used and where it may not, including prohibited use cases.
- Named ownership for AI risk, separate from but aligned with the CISO.
- A documented oversight mechanism for agentic systems who can authorise an agent, what guardrails constrain it, and who is accountable when it acts.
AI governance is not an innovation policy; it is a risk policy. The board should treat it accordingly.
Securing AI systems
AI introduces risks that traditional security was not built for: prompt injection, model exfiltration, data poisoning, and the leakage of sensitive information into prompts. Ask the security leadership where the organisation's models, prompts and training data live, who can access them, and what prevents an attacker from manipulating the model's behaviour.
The board does not need to understand the attacks in detail. It does need to ask whether the security function has been given the authority and budget to extend its controls to AI systems and to confirm the answer is yes.
Third-party and supplier AI risk
For most organisations, the most dangerous AI is not the one they built; it is the one embedded in a vendor's product, quietly processing their data. AI supply-chain risk includes models that change behaviour after an update, vendors that train on customer data, and agents that connect to third-party APIs with more privilege than anyone realised.
The board should ask which third-party AI systems touch the organisation's data or decisions, what the contracts say about model changes and data use, and whether the organisation can detect a vendor's AI behaving differently tomorrow than it did today.
Data exposure and confidentiality
Agentic systems are powerful precisely because they can reach data. That reach is also the risk. An agent with access to customer records, financial systems or trade secrets can expose them through a prompt, a tool call or a leaky integration. The board should ask: what data can our AI systems reach, and is that access necessary and monitored?
The principle is simple: give an agent the least access that lets it do its job, and no more. Most organisations cannot currently answer what their agents can reach. That is the first gap to close.
Oversight questions for the board
A short set of questions keeps AI governance on the agenda:
- Where is AI in use in our organisation, including through vendors?
- Which of those systems are agentic, and who authorised each?
- What is our risk appetite for autonomous decisions, and where is it documented?
- What guardrails and kill-switches constrain our agents?
- How would we know if an AI system caused harm and who would be accountable?
Governing AI well is not about slowing adoption; it is about making adoption defensible. Boards that build this discipline now will find agentic AI a manageable risk rather than a sudden one and a virtual cyber board advisor can help structure that oversight without slowing the organisation down.