How can a CISO govern generative AI?

A CISO governs generative AI by first finding what is in use, then giving every AI system an owner, a risk classification and a recorded approval with conditions. Those conditions are enforced at one control point that AI traffic passes through. Each decision and enforced request is recorded as it happens, so the organisation can show the board and auditors what it decided and what occurred.

The work is familiar. It is asset management, access control, data protection and third-party risk applied to a new kind of system. What is new is how fast use spreads, how much sensitive data travels in plain-language prompts, and how agents can act rather than only answer.

The five questions a board will ask

Boards rarely ask about models. They ask about exposure and accountability. Prepare an answer to each of these, with the record behind it.

  1. What AI do we use? A register of AI systems, both declared and observed in traffic, with vendor, models and purpose.
  2. What data goes to it? The data classes each system receives, and the rules applied to prompts before they leave.
  3. Who approved it? A named approver, who is not the requester, with conditions and an end date.
  4. What can it do? Whether it only answers, or whether it can read systems, send messages, change records or run commands.
  5. Can we prove it? Records made at the time of each decision and each enforced request, not assembled before an audit.

The risk picture and the control for each

RiskWhat it looks likePrimary control
Data exposure in promptsCustomer records, credentials or source code pasted into a chat or sent by an integrationPrompt inspection at a gateway with block, redact or log rules per data class, plus clear acceptable use rules
Shadow AITeams using AI services or local tools nobody has reviewedDiscovery from proxy and DNS logs and endpoints, then an intake route that turns findings into decisions
Agents and tool accessAn agent that can delete files, send email or call internal APIs on someone's behalfA register of agents and the tools they can call, least privilege, and a higher review tier for destructive tools
Prompt injectionText in a document or web page that changes what the model doesLimit what the system can do, keep a human in the loop for consequential actions, filter inputs where practical
Vendor and data processor riskA provider that retains prompts, trains on them or processes data in an unexpected regionVendor due diligence, contract terms, approved model lists and data residency rules at the gateway
Regulatory exposureUse cases that fall under the EU AI Act, GDPR or sector rules without anyone noticingAn inventory with purpose and affected people recorded, reviewed with privacy and legal
CostSpend spread across cards, teams and API keys with no ownerUsage and cost recorded per request, team budgets and one route for API traffic

Prompt injection is LLM01, the first entry in the OWASP 2025 Top 10 for LLM and generative AI applications. It is a design problem as much as a filtering problem, which is why limiting capability matters more than any single detection rule.

Who the CISO works with

FunctionOwnsThe CISO needs from them
DPOPrivacy reviews, DPIA where required, records of processingWhich personal data may be used, with which systems, under what conditions
LegalRegulatory scope, vendor contracts, terms of useThe interpretation of obligations and the contract terms that must hold
RiskThe classification method and risk appetiteWhat each tier requires and when a decision must go higher
ProcurementVendor onboarding and renewalsA trigger so no AI vendor is onboarded without review
FinanceBudgets and cost allocationWho pays for which usage and what limits apply
Business ownersPurpose, users, data and day-to-day use of each systemAn accurate record and acceptance of the conditions attached

The CISO typically owns security reviews, the control point where traffic passes, logging and incident handling. Keep the person who requests a system separate from the person who approves it.

A 90-day plan

Days 1 to 30: visibility

  1. Name the accountable executive and a small working group with privacy, legal and risk.
  2. Match proxy and DNS logs against known AI providers and list the results.
  3. Ask each team to declare the AI tools, integrations and agents it uses.
  4. Combine both into one register and assign a business owner to every entry.

Days 31 to 60: policy and approval route

  1. Publish a short acceptable use policy and a list of approved tools.
  2. Agree a classification method with a few factors, such as data sensitivity, autonomy and who is affected, and a written reason for each.
  3. Open one intake route with proportionate privacy, security and vendor reviews.
  4. Require an approver who is not the requester, conditions and an end date on every approval.

Days 61 to 90: enforcement and evidence

  1. Route API and browser AI traffic through one control point.
  2. Apply approved model lists, prompt data rules and team budgets there.
  3. Record each decision and each enforced or refused request at the time.
  4. Report the first metrics to the board and set a review cycle for approvals.

How to prepare for AI regulation

Regulation such as the EU AI Act assigns obligations by role and by the use a system is put to. Whether a system is in scope, and in which category, is a determination for your organisation and its counsel. The CISO's job is to make sure the facts for that determination exist and stay current.

  • Inventory. Every AI system with its purpose, users, data, vendor and the people it affects.
  • Classification by the organisation. A recorded classification decision for each system, made with legal and risk, with the reasoning kept.
  • AI literacy. Article 4 of the EU AI Act, as amended by the Digital Omnibus on AI, applies to providers and deployers. It asks for measures that support AI literacy among the people who operate or use AI systems on their behalf. Record who received what and why it fitted their role.
  • Records. Approvals, reviews, conditions and logs kept in a form an auditor can check.

Prohibited practices have applied since 2 February 2025 and Article 50 transparency obligations since 2 August 2026. The Digital Omnibus on AI moved high-risk obligations for Annex III systems to 2 December 2027. The EU AI Act checklist for deployers sets out the deployer duties step by step.

Metrics a CISO can report

Report counts and shares from your own records. Trends over time say more than any comparison with other organisations.

  • Number of AI systems recorded, split by declared and observed
  • Share of recorded systems with a named business owner
  • Share of systems with a completed classification and review
  • Number of approvals expiring in the next 90 days, and approvals awaiting review because facts changed
  • Observed AI services not yet in the register
  • Requests refused at the gateway, by reason, such as unapproved model, data rule or budget
  • Prompts redacted or blocked by data class
  • Agents with tools that can change or delete data, and how many have been reviewed
  • AI spend by team and model against budget

A CISO checklist

  • An accountable executive and a working group with privacy, legal and risk
  • A register of declared and observed AI systems, each with an owner
  • A short acceptable use policy and a list of approved tools
  • A classification method with readable reasons
  • Reviews and approvals with conditions, an end date and separation of duties
  • A register of agents and the tools they can call
  • One control point for AI traffic with model, data and budget rules
  • Records of decisions and enforced requests made at the time
  • AI literacy measures recorded
  • A small set of metrics reported on a fixed cycle

How MERIDVAR helps

MERIDVAR is self-hosted software that a CISO's team installs and operates on its own network, with no MERIDVAR cloud and no telemetry.

  • Visibility. An AI inventory of systems observed in traffic or declared by a person, and shadow AI discovery against your sanctioned provider list from proxy or DNS logs. Endpoint agents, with installers for Linux, macOS and Windows, report local AI tools.
  • Agents. An agent registry with Ed25519 signed identities and a declared MCP server inventory with tool-level risk.
  • Decisions. Rules-based risk classification from Low to Critical with a reason for every factor, plus privacy and security reviews. Approval is by an administrator who did not make the request, with conditions and an end date.
  • Enforcement. An admission gateway applies policy such as approved model lists, prompt DLP in ten categories of structured identifiers, data residency and team budgets. A refused request never reaches the model provider.
  • Evidence. An append-only, hash-chained evidence log and an audit package per AI system or vendor that states what it does not prove.

Controls apply only to traffic routed through the gateway, MERIDVAR does not certify compliance, and the product is in beta. See AI security for the details.

Sources: OWASP, 2025 Top 10 for LLMs and Gen AI Apps; OWASP, LLM01:2025 Prompt Injection; EU AI Act, Article 4 AI literacy; Regulation (EU) 2024/1689 on EUR-Lex; AI Act Service Desk, implementation timeline.