AI Governance Hub — The Ozkaya AI Governance Framework (AIGF)
Related AI and CISO resources: Continue with the AI Security Hub, Zero Trust Strategy Guide, Free CISO Toolkit, and Cybersecurity Leadership Brief.
AI Governance: Every CISO’s Most Urgent Unsolved Problem
AI is already reshaping the threat landscape — faster than most security programmes can adapt. The organisations winning at AI governance aren’t treating it as a separate workstream. They’ve integrated it into their existing security and risk frameworks.
Let me be honest about what I see in organisations today. The CISO is worried about ransomware and identity threats. The business is rolling out Copilot, ChatGPT Enterprise, and a dozen departmental AI tools — often without telling IT. And the legal team just found out about the EU AI Act. Nobody is coordinating.
This is the AI governance gap. It’s not theoretical — it’s happening right now in most organisations. IDC research from 2025 shows 65% of enterprise AI use is unsanctioned. That means sensitive data is entering external AI systems without data classification, consent analysis, or contractual protection. The organisation doesn’t know what’s being shared, with whom, or how it’s being used for model training.
At the same time, attackers are using AI to craft more convincing phishing, automate vulnerability discovery, accelerate malware development, and generate deepfake audio and video for social engineering. Cisco’s 2025 threat intelligence data shows AI-assisted attacks escalate 3× faster than traditional campaigns. The governance problem and the defensive problem are converging simultaneously.
Microsoft Digital Defense Report 2024
IDC Future of Work Survey 2025
PwC Global Digital Trust Insights 2026
Cisco Cybersecurity Readiness Index 2025
PwC Global Digital Trust Insights 2026
The Ozkaya AI Governance Framework (AIGF) organises AI governance into five domains that map directly to how security and risk teams already operate. It’s designed to be integrated into existing security programmes — not run as a separate initiative that nobody owns.
AI Inventory & Classification
Build a complete inventory of every AI system in use — sanctioned and shadow. Classify each by risk level: what data does it process? How autonomous are its decisions? What’s the business impact if it fails or is manipulated? You cannot govern what you cannot see, and most organisations cannot see most of their AI.
Data Governance for AI
AI systems are only as trustworthy as their training data and the data they process in operation. This domain covers data quality standards, consent for AI use, data residency requirements, and the critical question of whether sensitive data is flowing into external AI systems without adequate contractual and technical controls.
Model Security & Integrity
AI models face novel attack vectors that traditional security tools don’t address: adversarial inputs that manipulate outputs, model poisoning through corrupted training data, model inversion attacks that extract training data, and prompt injection attacks against LLM-based systems. The OWASP Top 10 for LLMs is your starting point.
Regulatory Compliance
EU AI Act risk classification is not optional for organisations operating in Europe. High-risk AI systems — covering employment decisions, credit scoring, critical infrastructure management, and law enforcement — require conformity assessments, transparency documentation, human oversight mechanisms, and registration in the EU database.
AI Incident Response
AI failures are different from traditional IT incidents. Model drift, bias amplification, and gradual output degradation can cause harm over weeks before anyone notices. AI incident response plans must include model monitoring thresholds, rollback procedures, harm assessment frameworks, and clear escalation paths for AI-related incidents.
The threat side of AI is moving faster than the governance side. Here’s what I’m seeing in the field and in the intelligence reports:
- AI-generated phishing — hyper-personalised spear phishing at scale, using scraped LinkedIn, news, and social media data. Click rates are 3-5× higher than generic phishing. Your awareness training hasn’t kept up.
- Deepfake voice and video — CEO fraud (Business Email Compromise) now includes deepfake audio calls impersonating executives. A UK engineering firm lost £20M in a single deepfake video call authorising a wire transfer.
- AI-assisted vulnerability discovery — attackers use AI to scan target environments, correlate exposures, and prioritise attack paths faster than any human team can patch.
- Prompt injection — attackers embed malicious instructions in content that gets processed by your LLM-based applications, causing the AI to exfiltrate data, bypass controls, or take unauthorised actions.
- Model poisoning in supply chain — compromising AI models during training or through fine-tuning data poisoning to introduce backdoors or biases in models you then deploy.
| AI Use Case | EU AI Act Risk | Key Requirement |
|---|---|---|
| HR recruitment & CV screening | HIGH RISK | Conformity assessment, human oversight, transparency |
| Credit scoring & loan decisions | HIGH RISK | Explainability, audit trail, bias testing |
| Critical infrastructure management | HIGH RISK | Full conformity assessment + registration |
| Customer service chatbot | LIMITED RISK | Disclosure that it is AI (transparency obligation) |
| Security anomaly detection | LIMITED RISK | Human review of high-impact decisions |
| Marketing personalisation | MINIMAL RISK | Good practice guidelines, voluntary code |
| Internal productivity tools (Copilot) | MINIMAL RISK | Data governance & acceptable use policy |
| Deepfake generation / manipulation | PROHIBITED | Banned outright under EU AI Act |
— Dr. Erdal Ozkaya, Author of 26 Cybersecurity Books & NATO Advisor
The EU AI Act entered into force August 2024. The prohibited AI systems ban is already in effect. High-risk system obligations apply from August 2026. Here’s what your action plan should look like:
- Complete your AI inventory — every system, every use case, every vendor. This is non-negotiable and should have started yesterday.
- Classify each system against the EU AI Act risk categories. Your legal team needs to be in this conversation.
- For high-risk systems: begin conformity assessment, technical documentation, and human oversight design now — you have less time than you think.
- For all systems: ensure your AI acceptable use policy is in place, communicated, and enforced. This is your first line of defence against shadow AI.
- Appoint an AI governance owner — this cannot sit solely in legal, IT, or security. It needs cross-functional ownership with clear accountability.
CISO Toolkit
AI governance templates, risk assessment frameworks, and board reporting guides for AI risk.
ISO 27001 Toolkit
Information security controls framework — the foundation for AI governance integration.
Sentinels Talk Show
AI security and governance conversations with CISOs, regulators, and AI practitioners.
26 Cybersecurity Books
Dr. Ozkaya’s full published library including titles on AI security and digital transformation risk.
Book Dr. Ozkaya
Board-level AI governance workshops, CISO advisory sessions, and keynotes on AI risk.
What your board needs to hear about AI governance
Three talking points, one metric, one question. Screenshot this for your next board prep.
AI governance is not a future problem. Your employees are already pasting customer data into public AI tools today. The first board question is not “what is our AI strategy” — it is “what is currently leaving the building.”
Treat AI risk as a data classification problem first, model risk problem second. If you cannot answer where your sensitive data flows, no governance framework on top will save you.
The real exposure is third-party AI inside SaaS tools you already pay for. Every vendor is shipping AI features by default. Your existing contracts almost certainly do not cover this.
AI Governance Is a Board-Level Problem. Let’s Solve It Together.
Most organisations are 12–18 months behind where they need to be on AI governance. The EU AI Act compliance clock is running. I work with executive teams to build AI governance programmes that are practical, proportionate, and actually implemented — not just documented.
Start the Conversation →
AI Governance FAQ — Honest Answers to the Questions CISOs Actually Ask
What is AI governance, and why has it become a CISO problem instead of just a compliance problem?
AI governance is the organizational framework — policy, technical controls, accountability structures, and audit trails — that determines how AI systems are selected, deployed, monitored, and retired safely. It became a CISO problem the moment generative AI moved from research labs to every employee’s browser. Three forces converged in 2024–2025: regulators shipped binding rules (EU AI Act, US Executive Order 14110, ISO/IEC 42001), enterprises woke up to data exfiltration through public AI tools, and boards started asking “who is accountable when our AI gives wrong advice?” The answer is the security function, because nobody else has the operational muscle for continuous risk monitoring. Compliance teams write the policy; CISOs make sure it actually holds in production.
How does the EU AI Act actually affect organizations outside the EU?
Same way GDPR did — extraterritorially. If your AI system affects EU residents, processes EU data, or is offered to EU customers, you’re in scope regardless of where you’re headquartered. The Act categorizes AI systems by risk: prohibited (social scoring, real-time biometric ID in public), high-risk (employment, credit, critical infrastructure, law enforcement), limited-risk (chatbots, deepfakes — transparency obligations), and minimal-risk (most internal tools). High-risk systems require risk management systems, data governance, technical documentation, human oversight, and registration in the EU database. Penalties hit up to €35 million or 7% of global annual turnover — higher than GDPR. If you’re a US, UK, or APAC company touching the EU market, treat this as binding now.
What are the highest-priority controls I should implement for safe enterprise AI use today?
Five controls that compound: (1) inventory every AI system in use, including shadow AI — most organizations don’t know what employees have signed up for, (2) classify by data sensitivity so you know which use cases can touch crown-jewel data and which can’t, (3) enforce data loss prevention at the prompt layer — block sensitive data from leaving your environment via AI APIs or browser plugins, (4) require human review for AI-generated decisions in regulated workflows (hiring, lending, medical, legal), and (5) instrument continuous monitoring for prompt injection, model drift, and output anomalies. The single highest-leverage control is the inventory — you can’t govern what you can’t see, and most enterprises are flying blind on this.
How is ISO/IEC 42001 different from the NIST AI Risk Management Framework, and which should I adopt?
Both, but they serve different purposes. ISO/IEC 42001 is a certifiable management system standard — it tells you how to structure organizational accountability for AI, similar to ISO 27001 for security. NIST AI RMF is a voluntary framework giving you the risk taxonomy, threat modeling approach, and lifecycle controls. Use NIST AI RMF for the practical risk work and ISO 42001 for the governance structure that survives audits. Federal contractors and regulated industries should plan to be ISO 42001 certifiable within 18–24 months. The mistake organizations make is picking one and treating it as comprehensive; they’re complementary, not competing.
How do I handle the “shadow AI” problem — employees using ChatGPT, Claude, Gemini, or Copilot on work data without approval?
Banning doesn’t work. Employees have already adopted these tools, often productively, and prohibition just drives use underground. The realistic playbook: (1) survey honestly to understand what’s actually being used and for what — most organizations are shocked, (2) provide sanctioned alternatives with enterprise data agreements, SSO, and audit logging, so the legitimate path is also the easy path, (3) implement DLP browser controls and CASB policies to block sensitive data leaving via unsanctioned AI tools, (4) update acceptable-use policies with specific examples — vague language fails, and (5) train continuously, not annually. Treat shadow AI like shadow IT was a decade ago: a signal of unmet legitimate need, not a discipline problem.
What should the board be asking the CISO about AI risk, and how should I answer?
Boards should be asking five questions: (1) what AI systems are in use across the enterprise, and who approved them? (2) what is our exposure if an AI system makes a materially wrong decision? (3) how are we complying with applicable regulations and what’s our remediation timeline? (4) how would we know if our AI was being manipulated, and how fast could we respond? (5) what’s our incident response plan when an AI-driven decision causes customer harm? The honest answer to question 1 in most enterprises today is “we don’t fully know” — and that’s the moment the conversation shifts from technology to governance. If your board isn’t asking these yet, brief them. They’ll be asking by the next meeting whether you raised it or not.