Latest / TeqTalk with Jas / AI Data Security Explained: Where Your Data Goes When AI Touches It
Transcript
- Ep71: Your AI might be secure. Can you tell me where your customer data goes after an AI agent touches it? Today that data can be retrieved, combined with information from another system, passed into another agent, written into an application, or used to trigger an action before anyone can see the final answer. IBM's 2026 cost of a data breach research found that more than 20% of organizations studied, reported a breach targeting AI models or applications. Most leaders still ask whether the model is secure, whether the data is encrypted, and whether the vendor is compliant. Those questions really matter, but they do not tell you how information moves once AI gets access to it. So the question, which is very important, is when AI touches a company data, where does it go? And can you really follow it all the way through? AI security is increasingly a data movement and action control problem now. In the next 10-15 minutes, I'm going to break That into five parts. What enters AI and what AI can do next? What happens when employees get involved? How to track the flow, and a four-question test that you can run against one of those AI workflows. And in each one of those five parts, I'll unpack the common AI security myth. Something that sounds right at first but doesn't hold up once agents, tools, memory, and enterprise data start working together. I am Jaskar, this is Tech Talk. Let's deep dive in. Part 1. Data in. What can AI reach? So the myth here is if our data isn't used to train a model, our data is safe. Well, training is only one of the ways AI get access to enterprise information. The fourth door is definitely training and fine-tuning. If you are tuning a model on internal data, you need to know exactly what went customer information, source code, support tickets, finance data, and whether that information was For that purpose. The second door is retrieval or RAG. And for many companies, this is more important. Imagine somebody asking an AI assistant, what is the risk of our customer renewal? In seconds, the system may retrieve CRM history, contracts, support cases, pricing, files, and internal notes. None of that information trains the model, but the AI can still bring it together. The risk is not just access to one document, it is The availability to combine information across the system. The third door is context and memory. System prompts, business rules, conversation history, agent straight, and persistent memory can all contain sensitive information. If an agent is told the renewal is worth $10 million, margin must stay above 30%, and discounts above 15% need CFA approval, that instruction itself is company data. The fourth door is The AI data layer, which is embeddings, vector databases, caches, intermediate artifacts, and when documents are indexed for AI retrieval, you have created another representation of enterprise informations that also needs governance. So in 2026, the better question isn't simply was my data used for training? It is how many places can my data become AI accessible? Part two is data out. What can AI do with it? Well, the myth here is if the agent is approved, the actions it takes are approved. Well, an approved agent can still be over permissioned. It may have access to tools, APIs, connectors, and systems that allow it to do far more than the original use case required. Well, today AI can do four things. It can generate generate a summary, proposal, report, or code. It can send information to Slack, email, API, or any of the applications. It can write, write into the system such as CRM ticketing databases, and lastly it can act by triggering workflows, executing codes, submitting transactions or requesting approvals. Here is where agent identity matters. With traditional Gen AI, we mostly ask what did this model say? Agents now we also need to know what identity performed the action and what authority came with it. An agent shouldn't inherit board permission simply because the person using it has them. So the risk expands further when agents call other agents. A sales agent may call pricing, pricing can call finance, finance can call approval. And one workflow can cross several trust boundaries before the user sees the result. So the important question is no longer just what the AI produced, it is What authority the agent had while producing it? It's so imperative nowadays. Part three is workforce. What happens when people get involved? Well, the myth here is if an AI tool is approved by the company, the employees can put company data into it. So this is where enterprise policy gets more complicated. Start with Shadow AI. Well, someone uploads a spreadsheet into an unapproved assistant, pastes the customer information into a chatbot. Or put source code into the tool that IT doesn't even know about it. But even approved AI doesn't mean approval use. An employee may have permission to open a confidential financial document inside a company, but that doesn't automatically mean the information should be entered into every model, every plugin, connector, or external services available inside an AI platform. Then there is derived data. A confidential spreadsheet becomes an AI summary. The summary becomes the Word document, the document becomes a slide, and the slide becomes the email attachment. So the format changed, the sensitivity did not, AI changes the representation of data, it doesn't erase its classification. And that's a workforce challenge. People tend to think about the file they started with, while AI systems can really create new versions, summary, and outputs continuously. The policy therefore has to follow the information. Not just the original file. Part 4 is control monitor track show. Myth here is if we log AI activity, we have traceability. Logging is necessary, but a collection of logs is not the same as bringing able to reconstruct what happened. One system knows a user logged in, another knows what the model received, another records a tool call, another knows which business record change. So technically everything may have been logged, but the story is fragmented. That is why I like three simple controls, which is monitor, track, and show. Monitor means knowing your AI surface, what models, agents, tools, and connectors exist, and which enterprise systems they can reach. Track means maintaining a classification and lineage as information moves. If a confidential data becomes an AI generated summary, the summary remains confidential. Lineage should let you move backward and identify the source and the Contribute to an answer and move forward to see whether that information really went afterward. Show means being able to reconstruct the transaction end to end. Who initiated it, what data was retrieved, which agent or model processed it, what tool was called, what changed, and where the result went. Observability tells you something happened. Traceability tells you how it happened. So if you cannot reconstruct the path, you cannot prove you are in control. Part 5 is the four question test. And myth here is human in loop means the workflow is governed. Human approval really helps, but only when the person has enough context to make a meaningful decision. Imagine an agent recommends a 12% discount and a manager clicks and approve. If the manager doesn't know which customer data pricing rules or the system actions produce that recommendation, the presence of the human doesn't automatically make the workflow governed. So pick one real AI interaction from last week, an employee using an assistant, a RAG application answering the question, or an agent performing an action, and then ask four questions. First, data. What sensitive information did the AI actually touch? Not our knowledge base, which document, records, customer information, context, memory, all of that. Second, govern. Was a person or agent allowed to use the information for this purpose and through this workflow? Think identity, data, purpose, and destination. Third, investigate. Can you trace the complete path from source to AI to tool or agent to action to destination? And can you do it quickly? Or would several teams need several days to rebuild the story? This is the one I would check first. Fourth, enable. If you can answer the first three with confidence, can you approve the next AI use case faster? That's the point of a good governance. It shouldn't slow AI down, it should make safe use cases easier to approve. Security should be an AI accelerator, not an AI bottleneck. And here was the point: know what data goes into AI, what authority AI has once it gets that data, how people transform and move the outputs, and whether you can trace the full path afterward. And then test one of the real interactions, data govern. investigate enable. The biggest challenge with AI isn't simply a new model. It is a software that can now interpret information and can take actions across several systems at machine speed. So do not just ask if my AI is secure. Ask can I follow my data wherever AI takes it. If the answer is I do not know that's where to start. Tell me in comments which of those four questions could your team answer today and which ones they cannot. I am Jazz Carr and this is Tech Talk. See you again next time. Thank you.