AI security is moving up the board agenda for good reason. A March 2026 analysis from Cycode identifies 11 critical AI security risks, and reports that 81% of organisations lack visibility into how AI is being used inside their business. That combination — growing risk and limited oversight — is exactly the point at which boards should ask harder questions.
The visibility gap
Most boards have a general sense that employees are using ChatGPT, Copilot or other AI tools. Fewer have a reliable inventory of which tools, which data is being shared, and whether any of it violates confidentiality, data protection or contractual obligations.
The 81% figure means that, in most organisations, AI use is at least partly invisible to security and compliance teams. Shadow AI is not just a buzzword; it is a measurable gap in risk management. Until you know what is being used, you cannot assess whether it is being used safely.
The 11 risks in plain language
Cycode groups the risks into familiar categories with AI-specific twists:
- Data exposure. Sensitive prompts, documents or code sent to third-party AI services.
- Prompt injection. Malicious inputs that trick a model into revealing data or taking unintended actions.
- Insecure model outputs. Over-reliance on AI-generated code, advice or decisions without verification.
- Supply chain risks. Vulnerable AI libraries, model weights or training pipelines.
- Excessive permissions. AI tools granted broader access than the user should have.
- Insecure plugins and integrations. Third-party extensions that broaden the attack surface.
- Model theft or extraction. Competitors or attackers reconstructing proprietary models.
- Training data poisoning. Malicious data introduced to bias or degrade model behaviour.
- Adversarial inputs. Inputs designed to evade detection or fool classification systems.
- Inadequate logging and monitoring. Failure to detect when AI systems misbehave.
- Lack of governance. No clear owner, policy or lifecycle management for AI systems.
The list is comprehensive, but it is not theoretical. Each risk maps to incidents that have already happened in organisations of various sizes.
What boards should ask in Q3
Boards do not need to become security engineers. They do need to make sure AI security is treated as enterprise risk, not just an IT issue. Useful questions include:
- Do we have an inventory of AI tools and use cases, including shadow use?
- Who is accountable for AI security and how often do they report to the board?
- Have we tested our AI systems for prompt injection, data leakage and adversarial behaviour?
- What contractual protections do we have with AI vendors, and do they cover data handling, sub-processors and incident response?
- Are our incident response and business continuity plans updated for AI-specific failures?
The bottom line
Cycode’s analysis is a useful prompt for boards to move AI security from a technical concern to a governance priority. The firms that get this right in 2026 will not be the ones with the most advanced AI; they will be the ones that know where their AI is, what it can access, and what to do when something goes wrong.