Louis Zezeran
5. nov. 2026
Your Best Data Makes AI Work. It’s Also Your Biggest AI Risk.
Every organisation is being told to adopt AI, and fast. Copilots, chat assistants, specialised tools for each department and now autonomous agents are all arriving at once. They have one thing in common. To deliver real value, they need your data, and not just any data. They need the best you have.
That’s the paradox at the heart of this episode of the Cybercast by NEVERHACK. Host Louis Zezeran is joined by Ronan Murphy, co-founder of Getvisibility and Chief Data Strategy Officer at Forcepoint, who has spent more than 20 years in the security industry. The two first met at the Nordic Baltic Security Summit in Tallinn, and this conversation picks up where Ronan’s summit talk left off: how do you protect the knowledge that makes your organisation valuable, while still getting the benefits AI promises?
It’s a short, focused episode of around half an hour, and it’s packed with practical thinking for anyone being asked to “just give the business AI”.
From Infrastructure to Knowledge: Security’s Newest Frontier
Ronan opens with a quick history of the security industry. When he started out, the job was protecting infrastructure. Then the focus shifted to applications. Those applications moved into the cloud, and as they did, identity became the new perimeter.
Now, he argues, we’re in the next shift. We’re protecting knowledge itself.
Ronan calls that knowledge your alpha: the crown jewels. What counts as alpha depends on who you are. It might be your financial modelling, your research and development documentation, your pricing files or your customer database. A manufacturer’s alpha is its intellectual property. A consumer business’s might be its customer data. And beyond the headline crown jewels, every department, from HR and finance to operations, sales and marketing, holds data that is valuable to the organisation in its own way. The people who understand that value best are the data owners in each department.
Why AI Changes the Risk
Here is where AI enters. Ronan describes it as an “insidious” technology, because it finds its way into organisations in every shape and form. Whatever the flavour, the rule is the same: to get the benefit, you have to share your alpha.
Done in a controlled way, the operational gains can be tremendous. But that alpha is scattered across file shares, Google, Microsoft, Oracle, Amazon, SharePoint, OneDrive and Box. For 25 years, the industry has built controls around all of it, at the infrastructure, application and identity layers. Now, Ronan observes, organisations are saying “carte blanche” and letting every kind of AI loose on it.
The questions that follow are basic but profound. When AI gets access to your crown jewels, where does the data go? Who can see it? Is it used for training? Do your competitors benefit? Can people inside your own organisation see things they shouldn’t?
On the training question, Ronan doesn’t mince words. Model providers improve by learning, and businesses need to ask whether they’re helping a provider get smarter in a way that ultimately serves their competitors, who are paying for the same intelligence.
Sovereign AI Doesn’t Make the Problem Go Away
Louis raises the response many organisations reach for: if we don’t trust the big AI providers, we’ll run our own models, on-premises or on our own cloud infrastructure. Problem solved?
Not even close, according to both. Louis points out that this is where the sovereignty discussion usually stops, when it’s really just the beginning. Ronan makes it concrete. Imagine your salary information, disciplinary records or other tightly restricted data finds its way into an internal copilot. Suddenly anyone in the company can ask what a colleague earns, or what’s in their disciplinary file. The AI never touched the outside world, and you still have a serious incident.
The lesson: where the model runs is only one question. What data it can reach is the bigger one.
Start With the Outcome, Then Draw Your Red Lines
So where do you begin? Ronan’s answer is a sequence of questions, and it starts with the business, not the technology.
- What do you want AI to achieve? Better R&D, operational efficiency, higher productivity, automated workflows, even just organising calendars. You need a clear objective first.
- Where are your red lines? Ronan calls this the “less sexy” question. What must never end up in an AI system? Employee salary data is his recurring example.
- What guardrails do you need? Which data assets and which users should, and shouldn’t, have access?
- Is the data fuelling your AI the right data? Do you know where it lives, and are the right controls around it?
That last question is getting harder by the day. Data is growing exponentially, and AI is adding to the pile.
“Confidential” Isn’t Specific Enough Anymore
One of the most useful insights in the episode is about the labels most organisations already use. Typical classification schemes, such as internal, external, private and confidential, are what Ronan calls a “flat” taxonomy. And for AI, flat isn’t good enough.
Consider three documents, all marked confidential. One is a salary file. One is a disciplinary record. The third is confidential research you actively want to use to train an AI model. The same label covers all three, but your AI policy needs to treat them completely differently. You need to know not just that something is confidential, but what kind of confidential it is.
How DSPM Does the Heavy Lifting
Classifying an entire organisation’s data sounds impossibly painful. Ronan’s point is that technology has caught up. Tools can now analyse a document’s contents using a mix of techniques, from natural language processing to pattern matching, and determine with high accuracy that a file isn’t just “confidential”, it’s a salary document, or a disciplinary record, or R&D.
The category for this is Data Security Posture Management (DSPM): understanding the security posture of your data at scale. With that granular view, organisations can decide whether data sits in the right place, whether it should be used for AI training, and whether it should be reachable through AI prompts.
Louis asks two practical questions. What about documents with several types of information? They can carry multiple tags, for example “confidential”, “personal data” and “regulated under GDPR”, and your policy can then say that anything tagged GDPR stays out of AI training repositories and away from AI access. Ronan adds that existing tags should be sanity-checked too, because in many cases past classifications were simply wrong.
And how does classification stay current? There’s an initial body of work to scan your repositories, but data is dynamic. If someone pastes a block of sensitive IP or personal data into an innocent marketing document, you want that change detected the moment the file is saved. Following data as it moves and changes is, in Ronan’s words, one of the key ingredients.
That monitoring can work with an agent on each device or agentless, connected to file repositories and collaboration tools. Ronan notes that, with the rise of shadow AI, endpoint agents are making a comeback.
Shadow AI: The New Top Data-Loss Concern
Ronan cites the World Economic Forum’s view that, for the first time, data loss through shadow AI has overtaken ransomware as the top data-loss concern. Employees are copying and pasting large amounts of content into AI tools their organisation never approved.
Crucially, he stresses that this usually isn’t malicious. People are chasing productivity and trying to be better at their jobs. But the result is the same: alpha leaking out to model providers.
Louis brings it home with his own experience. A company-approved assistant comes with reassurances about data protection, but if it isn’t capable enough, people are tempted to reach for something better, and end up making their own judgement calls about what’s safe to paste. Organisations need to give employees both clear controls and sanctioned tools that are genuinely good enough to use.
Delegating Risk to the People Who Understand It
Another standout idea: security teams shouldn’t have to be experts in every department’s data. When HR wants a new AI system, is it the security professional’s job to understand the risks in HR data? Arguably not. Instead, security can give HR visibility, dashboards and a view of risk, and let HR, as the data owner, decide what needs fixing.
Ronan goes further. There’s an old saying that the biggest enemy of security is compliance, and historically the two teams haven’t worked well together. AI is forcing them together, because everything now happens at machine speed. What’s emerging is a working triangle of security, governance, risk and compliance, and the data owners in each department. With agentic AI, which doesn’t just read data but edits, changes and shares it, that collaboration is becoming essential. In Ronan’s experience with large North American organisations, the CISO usually leads this, sometimes alongside the CIO, tackling one department at a time.
The Monday Morning Takeaway
Louis closes with the scenario many listeners will recognise: it’s Monday morning, there’s no budget, and the CEO just wants “the AI”. Ronan’s advice is to start from the outcome you want, then work backwards. Identify your crown jewels. Define your red lines: what must never reach a system that can share it with third parties. Get basic hygiene right. Do that, and you can accelerate AI initiatives with confidence and get the return on investment you expect.
Do it wrong, and after 20 years in the industry, Ronan believes it may be the single biggest threat to an organisation’s data security, governance and compliance.
Why This Episode Matters
AI adoption is being driven from the top, often faster than security, compliance and data owners can keep up. This episode gives you a clear framework for the conversation: know your alpha, move beyond flat classification labels, set red lines before you connect any AI, and share responsibility with the people who actually understand the data.
Listen to the full episode now, explore more at the NEVERHACK website, and subscribe to the NEVERHACK Cybercast so you never miss an episode.
