AI Security Has No Owner, and That Is the Problem
A sales team connects an AI assistant to the customer database to speed up account research. The data science team checks the model gives accurate answers. Legal reviews the supplier contract. Three months later, an internal audit finds the assistant will hand over any customer record to anyone who phrases the request the right way. The security team hears about the tool for the first time in the audit meeting.
Nobody in this story did anything wrong by the standards of their own job. The failure sat in the gap between them.
Everyone Uses AI and Nobody Owns the Risk
Pentera’s AI and Adversarial Testing Benchmark Report 2026 surveyed 300 CISOs and senior security leaders in the US. 67% said they had limited visibility into how AI is used across their organisation. Not one respondent claimed full visibility. The report describes ownership of AI systems as split across cloud, identity, application and data teams.
Security leaders are part of the pattern too. UpGuard’s State of Shadow AI research, published in November 2025, found 68% of the security leaders it surveyed had brought unauthorised AI tools into their daily work. If the people writing the policy work around the policy, the problem lies in structure, not discipline.
Two Professions Speaking Two Languages
I’ve spent over 17 years in cyber security, and I see the same divide in almost every organisation I work with. AI specialists talk about accuracy, training data, drift and performance. Security specialists talk about threats, vulnerabilities, controls and incidents. Both groups are good at their jobs. They rarely sit in the same meetings.
The newest risks fall straight into the gap between them:
- Prompt injection, where an attacker hides instructions inside content an AI tool reads
- Data poisoning, where tampered training or reference data changes how a model behaves
- AI agents given broad access to systems, with nobody reviewing what they are allowed to do
- Sensitive data pasted into public AI tools by staff trying to save time
A data scientist sees prompt injection as a quirk of model behaviour. A security analyst sees an input validation flaw. Each assumes the other team is dealing with it. Often, neither is.
Agents Raise the Stakes
AI agents make the gap more expensive. A chatbot answers questions. An agent takes actions. It sends emails, updates records, moves files and calls other systems on your behalf.
Every permission an agent receives becomes a route in for an attacker who manages to steer it. Access reviews were built for people and service accounts, not for software deciding its own next step. When the identity team, the AI team and the security team each assume someone else reviewed an agent’s permissions, the agent ends up with more access than any new employee would get on their first day.
What Convergence Looks Like in Practice
Convergence does not mean merging two departments or creating another committee. It means treating AI security as one discipline with one accountable owner. Here are 5 steps to take this quarter.
- Name one owner. Give a single named leader accountability for AI risk across the business. Shared ownership across 4 teams usually means no ownership at all.
- Put security in the room on day 1. Add a security reviewer to every AI project at the design stage, before a supplier is chosen or a tool is connected to live data.
- Cross-train both teams. Teach data scientists the basics of threat modelling. Teach security analysts how models are trained, tuned and deployed. A half-day session each way closes a surprising amount of the gap.
- Add AI to your asset inventory. List every model, AI tool and agent in use, the data each one reaches and who approved it. You cannot protect what you have not counted.
- Test AI systems the way attackers will. Run red team exercises against your AI tools, including prompt injection attempts. Bring both teams into the debrief.
Why I Founded the AICSA
For years I watched AI specialists and security specialists attend different conferences, join different professional bodies and read different research. The people building AI systems and the people defending them had nowhere to meet as peers.
In late 2025 I founded the AI and Cyber Security Association, the AICSA, as a global trade association for the convergence of AI and cyber security. The aim is simple. Bring both communities together, share what works and build a shared language before attackers widen the gap any further.
No single association fixes this alone. Convergence happens inside organisations, one project and one conversation at a time. What a professional community offers is a place for practitioners to compare notes and stop solving the same problems in isolation.
Start With One Conversation
Go back to the sales team and their AI assistant. One 30 minute conversation between the data science lead and a security analyst at the start of the project would have caught the problem. You do not need a new budget line to make it happen. You need to put the right two people in the same room, early, and make one person accountable for the outcome.
About the Author
Lisa Ventura MBE FCIIS is the Chief Executive and Founder of the AICSA, the AI and Cyber Security Association. She has worked in cyber security for over 17 years and is the author of Artificial Intelligence in Cyber Security, published by Kogan Page in 2026. Find out more about the AICSA at www.aisec.org.uk.
Sources
- Pentera, AI and Adversarial Testing Benchmark Report 2026: https://go.pentera.io/ai-security-exposure-survey-2026-report
- Coverage on The Hacker News: https://thehackernews.com/2026/03/ai-is-everywhere-but-cisos-are-still.html
- UpGuard, State of Shadow AI (November 2025): https://www.upguard.com/press/new-research-from-upguard-reveals-68-of-security-leaders-admit-to-unauthorized-ai-usage.md
About the expert
Lisa Ventura MBE FCIIS
Chief Executive and Founder, AI and Cyber Security Association
Lisa Ventura MBE FCIIS is the Chief Executive and Founder of the AICSA, the AI and Cyber Security Association. She has worked in cyber security for over 17 years and is the author of Artificial Intelligence in Cyber Security, published by Kogan Page in 2026. Find out more about the AICSA at www.aisec.org.uk.
Read the full transcript · 195 words
Hello, I’m Lisa Ventura, Chief Executive and Founder of the AI and Cyber Security Association, the AICSA.
My article for Experts Corner is called AI Security Has No Owner, and That Is the Problem.
AI tools are now in almost every part of the business. Sales, HR, finance and marketing all use them. When I ask organisations who is responsible for securing those tools, I usually get several different answers, and that tells me nobody is.
Research from Pentera this year found 67% of CISOs have limited visibility into how AI is used in their organisation. None of them said they had full visibility.
The problem sits between two professions. AI specialists focus on accuracy and performance. Security specialists focus on threats and controls. Risks like prompt injection and data poisoning fall into the gap, and each team assumes the other is handling them.
In the article, I set out 5 steps to close that gap this quarter, starting with naming one accountable owner for AI risk. I also explain why I founded the AICSA to bring these 2 communities together.
Have a read, and let me know who owns AI security in your organisation.