GenAI Still Lives in the Browser. Your Guardrails Should Too

|
November 18, 2026
Digital security illustration of a glowing shield inside a circular target, representing enterprise GenAI guardrails and data protection in the browser.

The emergence of agentic AI in the enterprise has been big news in the last year, and rightly so. That’s because agentic AI acts as well as advises, and the speed of operations is such that humans often cannot step in. Agents require broad permissions to function and often act without human review.

The good news is that enterprises are increasingly aware of the risks that can be posed through agentic AI. The bad news is that humans often use GenAI the same way that they always have, and that traffic is still going up. Some estimate that 38% of employed adults use chatbots for tasks at work1 while other sources put it higher still, which is understandable because even sources studying the AI space can mix findings about AI chatbots and AI agents.

The continued use of “prompt-response” or conversational AI highlights two issues that organizations today are concerned about:

  • Data Loss Prevention (DLP) in prompts and files
  • Use of unsanctioned tools, often via personal accounts

Ongoing Data Loss

One of the biggest overall concerns about the use of AI is the loss of sensitive data—almost always by workers that are simply trying to get their jobs done. Types of data shared include source code, customer data, and financial or other types of strategy documents, as well as HR data, confidential content, and credentials. The growing issues around DLP are partly due to the continued use of unsanctioned tools. According to the IBM Cost of a Data Breach Report 2026, data loss or compromise was reported in 49% of security incidents involving shadow AI.2

Use of Unsanctioned Tools – Shadow AI

Regarding usage of unauthorized GenAI services (“Shadow AI”), 67% of users are using non-corporate accounts on their corporate devices to access AI services, a slight decrease from the previous year. However, 45% of employees are now considered regular users of AI (authorized or not) on their corporate devices, up from 15% in the previous year.3 Given the publicity around the issues that can come from Shadow AI, you may wonder why this is still an issue, particularly because, according to an Okta study published in May 2026, 90% of executives are confident that they can see which AI tools are in use, and 95% assume that employees use AI responsibly.4 When the same study turned to employees, however, the findings are not reassuring. Fifty-two percent of knowledge workers use unsanctioned AI tools and 24% do so regularly. In the US, the use of unsanctioned tools rises to 67%.

AI Risk Starts Where the User Meets the Prompt Box in the Browser

Given all of the security tools in a typical company, it is interesting to consider why the issues of DLP and shadow AI not only continue to exist but are actually growing. The answer is simple—network, identity, and endpoint tools are blind to many events that happen in the browser, which is where people use AI. Identity providers may see a login to an approved AI service, but not to an unapproved one. Network tools see the domain but not the prompt. AI tool vendors can see only the tools that are sanctioned, not the ones that are not.

Even with the growth of agentic AI, most human AI use remains conversational, in chat tools opened in a browser. One reason for the continued popularity of these tools is that there is no agent or app to install; the benefits of AI are as close as a browser tab. Meanwhile, the use of personal accounts skirts identity and network controls and 67% of users reach AI through non-corporate accounts on corporate devices.5 People also choose personal accounts for convenience: 80% say it’s simply easier.6 In the meantime, data leaks continue through typing or pasting into a prompt box or uploading a file into a chat window. The browser can see that content in context, while network tools see only encrypted traffic after the data has already left.

Because most security stacks have no tool dedicated to watching the browser, DLP and shadow AI continue unimpeded. At the same time, organizations broadly acknowledge the value of AI. They don’t want to block it, because such edicts are easy to make and notoriously difficult to enforce. Instead, they want to protect their data and get visibility into AI use.

Get AI Guardrails That Are Easy to Implement and Actually Work

That’s where the Menlo IRIS Browser Extension comes in. Menlo IRIS turns the browsers your users already have into secure enterprise browsers, delivering protection, visibility, and control inside the browser session, which is exactly where people use AI. There’s no browser swap, no new infrastructure, and no need to reroute traffic through a proxy, so you can put enterprise browser security in place fast without a long project.

With Menlo IRIS, you can:

  • See which AI tools are in use. Centralized logging records browser events by user, site, and data type, so you can see which AI sites and tools employees reach, including the ones you haven’t sanctioned.
  • Steer users toward sanctioned AI. Domain- and category-based web governance lets you allow approved AI tools and restrict unapproved ones, without the all-or-nothing blocking that pushes people to workarounds.
  • Keep sensitive data out of prompts and uploads. In-session controls on copy/paste, file uploads and AI prompts stop sensitive data before it reaches an AI tool, using the same DLP dictionaries and inspection rules as the full Menlo platform.
  • Govern AI browser extensions. An enterprise-wide extension inventory, with up-to-the-minute risk scores from Menlo partner Spin.AI, lets you allow, block, or force-install extensions, including AI extensions that can read page data.
  • Prove that your policy is enforced. Insights and custom reports give compliance teams the data they’ve been missing for SOC 2, HIPAA and GDPR, audits, and keeps cyber insurers informed. Logs can be streamed to your SIEM.

With Menlo IRIS, you can say yes to AI, with a policy you can enforce. Instead of banning the tools that employees rely on, security teams can sign off on an acceptable use policy for GenAI because they can finally see and control what employees send to AI tools. Menlo IRIS provides practical, in-browser control over web-based AI use that works alongside your other AI governance efforts and complements your existing network and endpoint controls rather than replacing them. And because the same controls apply to any web destination, AI guardrails are just one part of the protection you get. Add-ons such as HEAT Shield AI, for intent-based protection from zero-day phishing, and Menlo File Security with Content Disarm and Reconstruction (CDR) for clean, fully usable files extend that protection as your needs grow.

GenAI still lives in the browser. With Menlo IRIS, your guardrails can live there too.

Learn more about Menlo IRIS

See how Menlo IRIS delivers right-sized browser security, in the Buyer’s Guide

‍

Sources
  1. Pew Research Center, Americans and AI: Chatbots, Smart Devices and Views on Impact, June 17, 2026.
  2. IBM, Cost of a Data Breach Report 2026: The AI Tipping Point.
  3. 2026 Verizon Data Breach Investigations Report.
  4. Okta, AI Agents at Work 2026: The Agentic Enterprise and Security.
  5. 2026 Verizon Data Breach Investigations Report.
  6. Okta, AI Agents at Work 2026: The Agentic Enterprise and Security.

Menlo Security

menlo security logo
linkedin logotwitter/x logoSocial share icon via eMail
See the Menlo Browser Security Platform in Action