We have all heard the old joke: cybersecurity becomes a boardroom priority right after a major breach.
For years, enterprises relied on three classic lines of defence:
- First line (operational management): Teams and engineers who run operations and execute controls
- Second line (risk and compliance): Oversight teams who check policies, monitor regulations and flag risks
- Third line (internal audit): Independent reviewers who assess the first two lines and report to leadership
The board sat above all of this as a high-level review panel. The CISO, meanwhile, spent much of their time translating technical jargon into business terms and hoping the risk slide did not lose the room.
That playbook is broken. With cloud expansion and AI reshaping operations overnight, security can no longer be an IT checkbox. Forward-looking security leaders are now elevating the board into a fourth line of defence: no longer a passive review panel, but an active strategic layer that defines risk appetite, treats AI innovation as a managed business strategy, and builds enterprise resilience, guided by clear CISO telemetry and risk metrics.
1. Accountability and Governance: The Regulatory Push
Boards are not taking on this role by choice. Regulators are driving it. Global frameworks such as the EU’s NIS2 and the SEC’s disclosure rules place explicit oversight duties on management bodies and boards, and in India, expectations from SEBI, RBI and DPDP-related mandates point the same way.
- Personal exposure for directors: Oversight failure increasingly carries personal consequences. D&O underwriters now look for evidence of active cyber governance, such as documented board review of AI deployment risks. CISOs need direct board access, documented escalation paths and formal indemnity structures to protect both the organisation and the security team.
- A visible audit trail: If the board is a line of defence, it must leave evidence. Minutes should record the questions asked, the assumptions challenged and the trade-offs approved, such as a formal sign-off to accept legacy software exposure during a cloud migration.
- Who audits the fourth line? Boards should run an annual self-assessment of their cyber and AI oversight capability, or commission an independent review.
Practical Board Mechanics: Operating Under Pressure
A fourth line cannot run on quarterly slide decks alone. It needs operational clarity.
- A written AI risk appetite statement: Define prohibited uses, low-risk automated workflows and high-impact scenarios that need human sign-off. For example, ban unmasked customer PII in public LLMs, allow automated drafting of internal knowledge-base articles, and require legal approval for automated contract evaluation.
- Quantified dashboards: Replace raw vulnerability counts with five or six financial and operational metrics, such as monetary loss exposure, time to detect and contain, and the percentage of AI tools inventoried with a named business owner.
- Board-level tabletop exercises: Test decisions under stress with realistic AI scenarios, such as a deepfake audio message from the “CEO” requesting an off-schedule wire transfer, or a customer support agent accidentally sharing internal pricing sheets externally.
- Pre-agreed decision rights: Decide in advance who can shut down the primary e-commerce gateway during a ransomware attack, and when regulators must be notified. These calls happen in hours, not quarterly cycles.
- Board capability and structure: Assess cyber and AI fluency across directors, and consider giving primary oversight to a dedicated Technology or Risk Committee.
AI Security: The Operational Reality
When boards ask about AI risk, a policy like “do not paste sensitive data into ChatGPT” barely scratches the surface. Models are nondeterministic and data changes constantly, so AI security is a continuous operational challenge.
To investigate incidents properly, logging the user prompt is not enough. Teams need the full context: system instructions, retrieved vector database chunks, agent tool calls and final outputs, all fed into the SIEM alongside identity and network logs. If a procurement agent starts issuing unauthorised purchase orders, a prompt-only log makes root-cause analysis impossible. You need to see the exact retrieved chunk that misled the model.
- Shadow AI and third-party models: Employees routinely use unsanctioned tools. Maintain a complete inventory of approved and legacy AI tools with named owners, and make sure vendor contracts cover training-data use and audit rights.
- The board as a target: Directors are prime targets for deepfake voice and video, spear-phishing and board-portal compromise. Simple controls help, such as mandatory out-of-band phone verification before any board-level wire transfer.
The Rule That Matters Most: AI Assists, a Person Decides
None of this monitoring matters without a clear answer to one question: who is accountable when the model gets it wrong?
The simplest policy that holds under pressure is that whatever the model produces, whether a contract summary, a code fix or a credit risk score, remains a recommendation until a human signs off. An AI tool can analyse loan data and suggest a credit rating, but an underwriter must review and approve the final decision. Skipping this step under deadline pressure erodes ownership.
- Least privilege for agents: Give an agent only the reach its task needs. An agent that summarises support tickets should have read-only access to ticket text, and be blocked from writing to external databases or sending emails.
Resilience and Continuity
The safest AI posture is one that can absorb disruption without halting the business.
- Concentration risk: Relying on one or two foundation model providers creates a single point of failure. If a major provider has a long outage or changes its API terms, operations should fail over to an open-weight model hosted on private infrastructure.
- Cyber insurance realities: Review policies for AI-specific exclusions, sub-limits and conditions, so the board does not assume coverage that will not exist in a crisis.
Where This Leaves Security Leaders
None of this requires abandoning what security teams already do well. It means pointing the same instincts — scepticism, audit trails, continuous monitoring — at a new kind of system that speaks plain language instead of code.
Governance keeps the wheels on while the organisation accelerates into AI. Monitoring tells you the wheels are still attached. Most companies invest heavily in the first and almost nothing in the second, and that imbalance is what becomes a headline a year or two later.
Log the full context. Treat AI infrastructure as the critical infrastructure it already is. Keep a person in the loop on anything that matters. Do these consistently, and the AI conversation gets a lot less frightening.
