Latest / Cybersecurity Business with Fexingo: Security Companies, Breaches, and Enterprise Defense / The Rise of Shadow AI in Enterprise Security
Transcript
- Lucas: We spend a lot of time talking about securing the front door of the enterprise — the firewalls, the identity providers, the zero-trust architectures. But there is a new kind of breach happening not through the back door, but through the side window that nobody knew existed until it was already open. Luna: You mean Shadow IT? Because we have covered that before with shadow apps and unauthorized software installs. Lucas: Shadow IT has been around since the days of USB drives and personal cloud storage. But what we are seeing now is distinct because it involves AI. I am talking about Shadow AI — employees bypassing corporate-approved tools to use unauthorized generative AI platforms for sensitive work. It is not just about using an unapproved app; it is about feeding proprietary data into models that have no governance, no logging, and no retention policy. Luna: So the risk isn't just that they are using a competitor's tool, but that the data itself is leaving the secure perimeter and potentially being ingested by public models. Lucas: Exactly. And the scale is different. In the past, if an employee wanted to move data out, they might copy a few thousand files to a USB drive or send them via a personal email account. With AI, an employee can paste entire sections of code, customer contracts, or financial projections into a chat interface in seconds. The velocity of data exfiltration has increased dramatically, and the average employee often does not even realize they are creating a compliance violation. Luna: That feels like a huge cultural shift. Are people doing this maliciously, or is it just convenience? Lucas: It is almost entirely convenience. The pressure to adopt AI is so high right now that employees are building their own workflows. They find a free tier of a large language model, they paste in a draft of a marketing campaign or a snippet of source code, and they get instant results. From their perspective, they are being efficient. From a security perspective, they are creating massive exposure vectors. We are seeing cases where junior developers are using unvetted AI coding assistants that inadvertently commit intellectual property to public repositories. Luna: And if those models are training on the input data, that intellectual property could end up being part of the foundation for a competitor's model later on. Lucas: Right. That is the nightmare scenario. If you are a pharmaceutical company and you paste clinical trial data into a public AI tool, and that tool uses your data to refine its next version, you have effectively leaked trade secrets to anyone who pays for that service. It is a loss of data sovereignty that traditional DLP tools struggle to catch because the traffic looks like normal web browsing. It is HTTPS encrypted, it goes to a legitimate domain, but the content inside is highly sensitive. Luna: How are CISOs even detecting this if the encryption masks the content? Lucas: They are starting to look at behavioral analytics and endpoint detection rather than just network inspection. Some organizations are deploying browser extensions or agent-based monitoring that can flag when certain keywords or document types are being pasted into known AI interfaces. It is less about blocking the site and more about understanding the pattern of interaction. If an employee is rapidly copying and pasting large blocks of text into a chat box, that is a signal worth investigating. Luna: Does this require a complete overhaul of the security stack, or can existing tools be adapted? Lucas: It requires a shift in mindset more than a complete tech overhaul. You do not necessarily need to buy a new product immediately, but you do need to update your policies. The old policy said 'do not install unauthorized software.' The new policy needs to say 'do not submit confidential data to unauthorized AI services.' That distinction is critical because it targets the data, not just the application. We are also seeing a rise in AI Governance platforms that specifically try to map out which AI models are being used across the organization and whether they meet your compliance standards. Luna: So it is about visibility first, then control. You cannot stop what you cannot see. Lucas: Precisely. And visibility includes understanding the supply chain of these AI tools. Just because an AI platform claims it does not train on your data does not mean it is true. There have been instances where major providers changed their terms of service retroactively, meaning data submitted under the old rules became fair game for training. Enterprises need to know exactly which models they are using and what the data handling agreements are. Luna: That sounds like a legal minefield as much as a technical one. How do you negotiate terms for thousands of individual users? Lucas: You do not. You centralize procurement. The most effective strategy we are seeing is mandating that all AI usage go through a single, vetted corporate instance. If you provide a fast, easy to use internal AI portal that integrates with your existing identity provider, employees are much less likely to seek out external alternatives. It is about removing the friction. If the approved tool is better than the shadow tool, people will use it. If it is clunky and slow, they will find another way. Luna: But what about the departments that operate outside of IT, like Marketing or HR? They might not have the same budget or approval process. Lucas: That is where cross-functional governance comes in. Security cannot be the only gatekeeper. You need representatives from Legal, Compliance, and Business Units involved in setting the AI policy. When Marketing understands why they cannot use a certain tool for campaign ideation, they are more likely to comply. It stops being a battle between Security and the business and becomes a shared responsibility. We are seeing some companies create an AI Council that meets monthly to review new tools and assess risks collaboratively. Luna: I like that. It moves away from the 'no' culture and towards a 'here is how we do it safely' culture. Lucas: Exactly. And education plays a huge role here. Many employees genuinely do not understand the implications of data ingestion. They think of AI as a search engine, not a data processor. Training needs to be specific: show them examples of what constitutes sensitive data in an AI context. Is a project name sensitive? Is a code snippet? Is a customer email address? Give them clear boundaries. Luna: Are there any notable breaches yet that highlight this specific risk? Lucas: While major headline-grabbing breaches solely attributed to Shadow AI are still relatively rare, we are seeing smaller incidents accumulating. There was a recent case involving a mid-sized fintech firm where an analyst used a public AI tool to summarize regulatory filings. The tool’s developer later sold aggregated insights from millions of such prompts to hedge funds. The fintech firm’s competitive analysis methodology was essentially exposed without their knowledge. It was not a hack; it was a misuse of a consumer tool. Luna: That is terrifying. The data was used against them indirectly, and they never even knew it happened. Lucas: It underscores the importance of data classification. If you know what your crown jewels are, you can build controls around them. Tagging documents as Confidential or Restricted allows DLP systems to block those specific files from being uploaded to any non-corporate AI endpoint. It is a targeted approach rather than a blanket ban. Luna: So the solution is a combination of technology, policy, and culture. Technology to detect, policy to define, and culture to encourage adoption of safe tools. Lucas: Yes. And it is an ongoing process. As AI models evolve, so do the attack vectors. New features like voice input or image upload create new channels for data leakage. Security teams need to stay ahead of these capabilities. It is not a set-it-and-forget-it problem. Luna: Do you think we will see stricter regulations coming down the pipe regarding employee use of AI? Lucas: Almost certainly. Regulators are looking at how data is handled in AI systems, and employee misuse is a prime target for scrutiny. If a company can prove it had reasonable controls in place to prevent data leakage, it mitigates liability. If they turn a blind eye to Shadow AI, they are vulnerable. Expect gdpr style penalties to be applied to AI governance failures in the near future. Luna: It makes sense. Accountability follows the data. Lucas: Always. So for CISOs listening, start by mapping your AI usage. Talk to your business units. Understand where the data is flowing. And remember that trust is good, but verification is better. You need to see what is happening in real-time. Luna: Before we wrap up, I want to take a moment to thank our listeners. If these conversations help you navigate the complexities of business and technology, your support keeps this show ad-free and independent. You can support us directly at buy me a coffee dot com slash fexingo. It helps us continue producing deep-dive episodes like this one. Lucas: Thank you for that. It really does make a difference in maintaining the quality and frequency of our coverage. Now, let’s bring it back to the practical side of implementation. What is the very first step a small team should take if they suspect Shadow AI is already active in their organization? Luna: I would say communication. Before deploying any monitoring tools, announce the new policy clearly. Explain why it matters. Frame it as protecting the company’s assets rather than policing individual behavior. Transparency builds cooperation. Lucas: Agreed. And pair that communication with a quick audit of your current SaaS subscriptions. You might be surprised to find that several departments are already paying for AI tools out of their own budgets. Bringing those into the fold under a corporate agreement is often easier than trying to block them outright. Luna: That is a smart pivot. Instead of fighting the tide, ride it with guardrails. So, the key takeaway is that Shadow AI is not just a tech problem; it is a governance and cultural challenge that requires a holistic approach. Lucas: Spot on. It is about balancing innovation with integrity. As AI becomes more embedded in our daily work, the line between productivity and risk will continue to blur. The organizations that thrive will be those that adapt their security models to reflect that reality. Luna: Thanks for joining us for this deep dive into Shadow AI. Let us know how your organization is handling these challenges. Lucas: We will leave you with this question: In your experience, is the biggest barrier to adopting safe AI practices technical limitations, or organizational resistance?