In this article
"We locked down ChatGPT Enterprise and Copilot. So why is our security team still worried about AI data leaking?" Because in many companies the most exposed layer is not the AI tool you licensed. It is the browser extension an employee installed next to it, promising faster replies, a sidebar assistant, or a free VPN, with permission to read every page they open.
The direct answer: AI browser extension security is about controlling software that runs above your sanctioned AI tools. An extension with broad site access can read prompts, responses, internal URLs, and session data in any tab, regardless of what your AI vendor contract says. The fix is browser-level control, with an allowlist and permission-based blocking, not a reminder email.
What makes AI browser extensions riskier than the AI tools themselves?
An extension operates at the browser level, not inside one application. If it holds permission to read and change data on all websites, it can see everything rendered in the browser: the ChatGPT conversation, the HR portal, the finance dashboard, and the internal wiki. Your enterprise agreement with an AI vendor governs how that vendor handles your data. It says nothing about a third-party extension reading the same screen.
That is why a company can negotiate a strong AI contract, train staff on what not to paste, and still lose sensitive content the moment someone installs a free "AI assistant" add-on. The extension becomes an uncontrolled layer on top of a controlled tool. We have written before about why business owners should be careful when choosing extensions; the AI wave has made that warning far more specific.
How are attackers exploiting AI extensions?
The pattern is now well documented. In January 2026, SecurityWeek reported OX Security's finding that two Chrome extensions impersonating a legitimate AI sidebar tool had gathered more than 900,000 downloads while exfiltrating users' complete ChatGPT and DeepSeek conversations, along with URLs from every open tab, search queries, and URL parameters containing session tokens. The extensions asked users to consent to "anonymous, non-identifiable analytics" and then took the conversations instead.
Microsoft's own investigation adds the enterprise dimension. In its March 2026 analysis of malicious AI assistant extensions that harvest LLM chat histories, the Microsoft Defender Security Research Team reported roughly 900,000 installs and confirmed activity across more than 20,000 enterprise tenants. The extensions logged nearly every visited URL, including internal sites, plus excerpts of AI prompts and responses, and sent them to attacker-controlled domains at regular intervals. Telemetry was switched back on after updates even for users who had previously declined it.
Then there is the extension that did not look malicious at all. Koi Security found that Urban VPN Proxy and three related extensions, with more than 8 million installs between them, were capturing chatbot conversations across platforms including ChatGPT, Claude, Gemini, and Microsoft Copilot. As The Register reported, collection was enabled by default with no user-facing way to turn it off, it ran even when the VPN was not active, and Urban VPN carried a Featured badge from the Chrome Web Store.
What is actually at stake in a stolen AI conversation?
A stolen AI chat history is rarely trivial. Employees use these tools to draft client proposals, summarize contracts, debug code, and rework HR documents, so a captured conversation can hold exactly the material a confidentiality clause or privacy obligation is meant to protect. Microsoft's Digital Defense Report 2026 puts it plainly: LLM conversation history now routinely holds source code, architecture, customer data, and leaked credentials.
The business consequence is that an extension compromise can become a reportable data incident without any system you monitor being touched. There is no failed login to alert on and no unusual download from your file server. The data leaves through a browser extension nobody ever reviewed, and the first sign is often a vendor advisory naming an extension your staff already has installed.
Why store ratings and badges are not a trust signal
The common belief is that a popular, well-rated, store-approved extension is safe enough. The record says otherwise. A Featured badge did not stop Urban VPN from collecting AI chats. One of the AI sidebar extensions OX Security flagged had more than 300,000 installs and a 4.6-star rating before Google removed it, and in August 2026 SecurityWeek reported that the same extension was back in the Chrome Web Store, pushing a new malicious payload to enterprise endpoints through Google's own distribution infrastructure.
The deeper problem is that an extension is not a fixed product. It updates silently. The Urban VPN consent prompt arrived through a silent update, and the returning sidebar extension shipped a clean version first and added its payload in a later one. An extension that was reasonable when someone approved it can become something else a few weeks later. That is why approving extensions once, by reputation, does not work. The control has to be about what an extension is allowed to do, enforced continuously.
Why do extension permissions matter more than the listing?
Permissions tell you what an extension can actually reach; the listing tells you what it wants you to believe. Palo Alto Networks' Unit 42 found 18 high-risk AI browser extensions marketed as productivity tools, including one that requested debugger, all-URLs, web-request, and scripting permissions and could execute remote commands inside authenticated sessions such as email or a corporate VPN portal. Another read OpenAI, Gemini, and Claude API keys from browser storage and forwarded them to its developer. None of that is visible from the store description.
This connects directly to a question we raised about whether AI plugins were ever scanned before installation. Extensions are the same problem in a more familiar wrapper, and they rarely go through the vendor review a new SaaS contract would. Nobody submits a free sidebar for procurement, so it bypasses the controls designed to catch an overreaching permission request.
What should IT check before approving an AI extension?
Before any AI-related extension is approved for company devices, a short review catches most of the obvious risk. It should be fast enough that people use the process instead of working around it.
- Requested permissions. Flag read-and-change access to all sites, debugger, cookies, web request interception, and scripting. Ask why an AI writing helper needs any of them.
- Publisher identity and history. Confirm who publishes it, how long it has existed, and whether ownership changed recently. Sold or abandoned extensions are a common route for injected code.
- Where data goes. Determine whether it sends content to its own servers or only to the sanctioned AI vendor's domain.
- Threat reporting. Search the extension name and ID in security vendor reporting, including names it used previously.
- Business need. Confirm the sanctioned AI tool cannot already do the job. Often it can.
How do you enforce extension controls instead of relying on policy?
The enforcement lives in browser management. Google's guide to managing extensions in the enterprise lays out the approach for Chrome: decide which extensions to allow based on the permissions they request, block extensions based on those permissions, and manage allow and block lists through the Google Admin console or Windows Group Policy. Microsoft Edge offers equivalent policy controls, and Microsoft's own guidance in its extension investigation was to inventory, audit, and restrict browser extensions across the organization.
In practice, the model that works for a 20 to 300 person business is deny by default. Block all extensions except an approved list, block high-risk permissions even for approved extensions where possible, and force-install the few extensions the business genuinely needs. Then review the allowlist on a schedule, because approved extensions update. This complements the broader balance between browser accessibility and protection most teams already manage, and it extends naturally to the newer risk of agentic AI browsers already on your team's devices.
Is a written policy enough, or does this need training?
A written policy is necessary, but on its own it will not stop the next install. Employees need to understand why an AI extension is different from a theme or a spell checker: it sits where they type their most sensitive work. That understanding is what makes them flag a suspicious add-on instead of installing it, especially on personal or unmanaged devices where your browser policies do not reach.
Training should also reinforce the data side. Even a legitimate extension sees what users type, so the rules on what should never go into a general AI tool apply to anything running in the browser. And because extensions arrive the same way other unsanctioned software does, they belong in the same conversation as shadow IT inside your business.
How Securafy helps close the extension gap
When we work on this with a client, the first step is visibility: an inventory of every extension installed across managed Chrome and Edge browsers, matched against permissions and known threat reporting. Most organizations are surprised by the count and by how many extensions hold all-sites access. We then move the environment to an allowlist model through browser management, block high-risk permissions, remove anything without a clear owner or business purpose, and set a review cadence so approved extensions are rechecked after updates.
We pair that with endpoint and network monitoring so that if an extension starts calling unfamiliar domains, it is noticed quickly rather than discovered months later in a vendor report.
Where To Go From Here
If nobody has reviewed the AI extensions installed across your company's browsers, that inventory is the first step, and it usually takes far less time than people expect. Once the controls are in place, the remaining risk is human judgment.
If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. AI University gives your people role-based AI training with security built into the material, not bolted on afterward.
If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.
Know what a new tool can touch — use the Workflow Security Checklist before you connect anything to real data.
Get the checklist
Rodney Hall is the President and COO of Securafy, with 2 decades of experience in IT service management and operations.
He writes about the less glamorous but essential side of IT: support systems, documentation, business continuity, recurring issues, downtime, and the processes that keep client environments running well. His perspective comes from years spent improving how service is delivered, how teams respond, and how small problems are prevented from becoming much larger ones.
Outside of work, Rodney enjoys home improvement projects, woodworking, and dirt bike riding. His personal mission mirrors Securafy’s: helping businesses stay secure, compliant, and ready for whatever comes next.
Writes about: Managed IT, IT operations, service delivery, business continuity, downtime prevention, support processes, operational risk
Join the conversation
Have a question or a different take on this? Add it below.