Prevent data exposure in AI

In spring 2023, an employee from a household name electronics company leaked confidential information by using chat GPT to review internal code and docs. That incident led to a ban on Genai tools by that company and many enterprises followed their example. A few years on, a blanket ban on Gen AI apps may feel too extreme because the productivity and competitive advantages of using AI are too big to ignore. But every business is still worried about making headlines by exposing sensitive data with AI. Today, teams need a data protection strategy that moves beyond the binary decisions to block or allow and instead extend visibility controls over user interactions based on context. Cloudflare helps you do just that. Safeguarding data flows when your users interact with AI tools so you can safely adopt new tools with confidence. When a user inputs a prompt, Cloudflare detects, blocks, and logs that prompt with sensitive data. We do this with data loss prevention or DLP detections for PII, source code, customer data, financial information, credentials, and more. We recognize that traditional DLPs use simple keyword matching or regular expressions, which is insufficient for the nuance of AI traffic. That's why our DLP uses AI for AI security. We leverage multiple language models to understand not just the content but also the intent of prompts. This enables our AI powered prompt protection which includes systematic detections, topic classification and security guard rails for major platforms like Google Gemini, Chat GPT, Claude and Perplexity. So, if the topic of a prompt is intended to request PII or malicious code or attempts to get around an AI model's policies, Cloudflare will block and log that, too. Let's take a look at these data protections and topical guardrails in action. In this video, we're going to cover how Cloudflare Sassy solution protects sensitive data in your organization now that AI is a part of your employees workflows. We'll approach it from three places. Content detection, intent detection, and full visibility into what prompts people are sending. Let's start in DLP profiles. These are the detection rules we're going to apply to both data in motion and data at rest across your organization. If we look at a profile targeting personal information, we can see a number of predefined detections like bank account numbers and social security numbers. But moving over to another profile specifically designed for AI prompts, you'll notice it distinguishes between content and intent. and we'll dig into that difference in a minute. Below that, we're going to turn on AI context analysis. Now, this lets the DLP engine adjust its confidence in a detection based on the surrounding context. So, ultimately, you get fewer false positives for data at rest. Let's look at our CASBY integrations for enterprise AI services like OpenAI and Enthropic. We can see some of the major detections here are linked to DLP profiles that you apply to the integration. In the next page, we can see we've already found a lot of sensitive information that hasn't been properly stored. As I click into one of the findings now, if you're going to secure data in motion, you'll need to apply these DLP profiles to a secure web gateway policy as we're doing here. According to our expressions, if traffic matches one of the DLP profiles and if it's inside a Gemini prompt or upload, then we're going to block the request, log the DLP payload, and record the prompt for future reference. We'll also give the user a helpful prompt to coach them about proper AI usage policies. Now, I'm going to upload a bunch of customer data to the AI and ask it to parse it for me. Gemini is going to think for a while, but then eventually it returns a network error. Now, this is because my traffic is being on-ramp to Cloudflare's network before it reaches Gemini and it's applying the security profiles I set against my outbound traffic, which stops the message entirely before it leaves the Cloudflare network. You can also see a helpful coaching message appear at the bottom of the screen from our device client. Coming back briefly to the DLP profile, I'm going to disable the content filter on this page and simply leave intent. This means the AI won't do anything about sensitive information unless it detects the prompt is inappropriate, risky, or malicious. Coming back to the AI, we can see it parses the customer data and flags it for duplicates. However, when I ask the AI to tell me specifics about a high-v value customer, the prompt is considered to have malicious intent and stop the same way as before. Finally, in logs, I can see a list of all the AI related DLP violations that have occurred in my network. After decryting the log, I can see what my users have been asking and whether or not that poses any risk to my organization. The reality is that your workforce will likely overshare data with AI. In fact, a recent survey found that 93% of employees admit to putting information into AI tools without approval. Security teams must strike a balance between enforcing controls and unlocking productivity with AI usage. This demo focused on how Cloudflare can help you navigate that balance primarily with inline data protections. But those data protections also extend to outofband integrations via APIs or event hooks with leading gen AI tools like Chat GPT, Gemini, and Cloud. And with Cloudflare's broader AI security suite, you can secure the life cycle of AI communication across both humanto AI and machine to machine interactions. To learn more how Cloudflare secures AI adoption, check out the rest of our demo videos or try it for yourself by creating a free account on cloudflare.com.