An extension allowed to read and change data on all websites sees everything the employee sees in the browser: email, CRM records, internal dashboards, session cookies and full AI chat transcripts. That access is granted at install, by design, and Verizon’s 2026 Data Breach Investigations Report found the average company had more than 15% of users with unauthorized AI extensions. Two fake AI sidebars with over 900,000 installs scraped ChatGPT and DeepSeek chats straight off the page, security vendor OX Security disclosed in December 2025. Network tools rarely notice, because the stolen data leaves as ordinary HTTPS from the user’s own session. Most of the fix is browser management. Inventory every browser and profile, block extensions by permission and keep them off sensitive domains, then run a default-deny allowlist for admins, finance, HR and production engineers. Browser policy can’t judge what employees choose to paste or upload into AI tools, and it stops at the browsers you can put policy on. An endpoint agent that sees those actions in any app closes that gap. On managed Windows and Mac laptops, Jazz does this at the operating-system level, with no browser extension of its own, and investigates each movement.
Updated October 23, 2026.
Which tools can stop extension leaks, and what each one sees
Extension leaks have two parts that need different controls: what an extension reads on its own, and what employees deliberately send to AI tools. Each approach below covers some of that and misses the rest.
| Approach | What it sees for this problem | What it misses |
|---|---|---|
Browser management policy (Chrome and Edge ExtensionSettings, Firefox policies, Apple declarative settings for Safari) |
Which extensions can install, which permissions they may hold, and which sensitive sites they’re kept off | Browsers without policy, such as Arc and portable browsers; a malicious update that keeps the old permissions |
| Extension inventory (osquery and similar) | Every installed extension per browser and profile, with version, author and, for Chromium browsers, permissions | It blocks nothing; it tells you what to fix |
| Extension request review | Who asked for each extension, why, and with which permissions | Changes a publisher ships between reviews |
| Secure web gateway or CASB | HTTPS leaving the browser | An extension reading the page inside the user’s session and posting out as ordinary HTTPS |
| EDR | The browser process, plus known-bad extension IDs and domains once published | Whether a legitimate AI helper reading a customer record is acceptable use |
| DLP delivered as a browser extension | Activity in the browsers and profiles where it’s installed | Other browsers, personal profiles and private windows; it’s also one more extension with page-wide access |
| Endpoint DLP agent at the OS level, such as Jazz | On managed Windows and Mac laptops, what employees copy, paste or upload into AI tools in any browser or desktop app, and whether the destination account is personal or corporate | An extension silently scraping the DOM in the background; phones, Linux and unmanaged devices |
Browser policy and inventory are the controls for the extensions themselves. For a company whose people use Arc, Firefox or a portable browser alongside Chrome, and whose risk is customer data pasted into personal AI accounts, Jazz fits because its agent sees the paste at the operating-system level whichever browser is open, and Melody, Jazz’s agentic investigator, weighs whether that data should have gone there.
What can a browser extension actually read?
An extension reads whatever its permissions cover, and the broad ones cover nearly everything. Mozilla’s WebExtensions documentation covers the cross-browser extension model that Chrome, Edge and Brave also follow, and lists what each grant does:
- Host permissions decide which sites the extension can touch. The special pattern
<all_urls>matches every URL on http, https and several other schemes. On a matched site the extension can inject scripts, watch requests through thewebRequestAPI and, with thecookiespermission, read that site’s cookies. - Content scripts run inside the page. Per MDN, they “can access and modify the page’s DOM, just like normal page scripts can.” The DOM is the rendered page: the open email thread, the customer record, the text in a half-written form.
cookiesreturns the session cookies that keep the user signed in to each SaaS app. Google’s Chrome cookies API reference (updated September 2026) shows anhttpOnlyflag on each cookie it returns, so cookies hidden from the page’s own scripts are visible to the extension.clipboardReadlets the extension read whatever the user last copied.webRequestshows the extension every request the browser makes to matched hosts. Manifest V3 removed the ability to block requests for most extensions, but Google’s Chrome for Developers reference (updated September 2026) says the rest of the API “is unchanged and available for normal use.”tabsexposes the URL and title of every open tab, andscriptinglets the extension run code on demand in any page it has host access to.
AI extensions routinely ask to “read and change all your data on the websites you visit,” according to a Cloud Security Alliance research note from 11 April 2026. In practice that grant puts a third party inside every SaaS tab the employee opens, signed in as the employee, with no single sign-on prompt and no audit log in those apps. A companion CSA note from 10 April calls this “real-time read access to everything the user can see,” obtained through a consent dialog rather than an exploit.
The requests are common. Incogni’s January 2026 study of 442 AI-powered Chrome extensions found 42% requested scripting, a permission it says could affect about 92 million users, and 52% declared that they collect some user data.
Are AI browser extensions that read every page a data leak risk?
Yes. The documented cases fall into two groups that need different controls: fake lookalikes built to steal, and legitimate assistants that take more than users expect. Trusted extensions hijacked through an update are a third route, and AI assistants were among the victims there too.
Malicious lookalikes: fake AI sidebars
The largest documented case is a pair of extensions copying AITOPIA’s AI sidebar. “Chat GPT for Chrome with GPT-5, Claude Sonnet & DeepSeek AI” had over 600,000 users and Google’s “Featured” badge; “AI Sidebar with Deepseek, ChatGPT, Claude and more” had over 300,000. OX Security, which sells application-security products, reported them to Google on 29 December 2025. Its vendor analysis, covered by SecurityWeek and The Hacker News in January 2026, found the code looked for chat elements in the page on ChatGPT and DeepSeek and lifted the messages out of the DOM. Every 30 minutes it sent them, along with the URLs of all open tabs, to an attacker server. Users had agreed to share “anonymous, non-identifiable analytics data.”
Microsoft’s Defender research team followed up on 5 March 2026, saying its telemetry showed activity “across more than 20,000 enterprise tenants” (eWeek). The 900,000 install figure in that post is OX Security’s count, and Microsoft did not say how much data left those tenants.
A second campaign, which LayerX named AiFrame in February 2026, used around 30 extensions posing as AI assistants. LayerX, a browser-security vendor now owned by Akamai, counted more than 260,000 users; BleepingComputer reported over 300,000. The extensions loaded their interface from a remote server in a full-screen frame, so the operators could change behaviour without a store update. Fifteen of them read visible Gmail text, drafts included, straight from the page. A Google spokesperson later confirmed to The Register that all of them had been removed (story updated 17 February).
Legitimate AI assistants that collect more than users expect
Real products can take more than the user meant to share. A UC Davis and UCL team audited nine GenAI browser assistants for a USENIX Security 2025 paper and found they “often collect and share full webpage content, including the HTML DOM and user form inputs in some cases,” with their own servers. Merlin was the only one that recorded form contents. In the test it captured a Social Security number typed into an IRS refund form, plus records from a university health portal. HARPA collected the full DOM on all 20 test pages when it was used. Sider and the Copilot extension the team tested ran automatically when the user searched, and Monica and Sider built profiles from all five personal attributes the team measured. These were Chrome versions tested in 2025, and the products may have changed since.
Incogni’s 2026 study ranked Grammarly and QuillBot as the most potentially privacy-damaging AI extensions with over 2 million downloads, because both declare collecting website content and personal communications and need scripting and activeTab. Incogni scored both low on likelihood of malicious use, and its collection data comes from what developers declare on the Chrome Web Store (Help Net Security, January 2026). Incogni reported no misuse by either product.
Verizon’s 2026 DBIR puts the enterprise consequence plainly. One main function of AI plugins “is often to collect and retain information about what the user is browsing for context,” so if staff browse internal sites, “some of your non-public data might be getting vacuumed up.”
What happened to the Cyberhaven security extension V3, and how do trusted extensions turn malicious?
“Cyberhaven security extension V3” is the Chrome Web Store name of the browser extension sold by Cyberhaven, a DLP vendor. On Christmas Eve 2024 an attacker sent its developers a phishing email posing as Chrome Web Store policy enforcement. One developer approved a malicious OAuth app called “Privacy Policy Extension,” which gave the attacker the right to publish new versions. Version 24.10.4 was live for 31 hours, from 1:32 UTC on 25 December to 2:50 UTC on 26 December, and Chrome browsers running the extension in that window updated automatically (Ars Technica, January 2025). The listing had about 400,000 corporate users (TechCrunch). How many received the bad version was never published. The recovered payload went after Facebook cookies and access tokens, and Cyberhaven told customers that “authenticated sessions and cookies” could have been sent to the attacker’s domain.
It was one of many. Singapore’s Cyber Security Agency listed 25 extensions confirmed to carry malicious code as of 30 December 2024, including several AI assistants. Researchers’ counts ran from about 20 extensions with 1.46 million downloads to at least 33 on roughly 2.6 million devices (Ars Technica, January 2025), the higher count folding in a separate, older campaign. The earliest compromise dated to May 2024.
ShadyPanda shows the slow version. In vendor research published in December 2025, Koi Security, which sells extension security, reported that five extensions, three of them legitimate for years, pushed a malicious update in mid-2024 that fetched and ran new JavaScript every hour and sent every visited URL to the operator (The Hacker News). One of them, Clean Master, had been featured and verified by Google. Those five had about 300,000 installs. Another five Edge extensions from the same actor, with about 4 million installs, collected URLs, searches and clicks. The 4.3 million headline adds the two groups together.
A publisher can also change course without any attacker. The CSA’s April 2026 note records that Urban VPN Proxy, with more than 6 million Chrome and 1.3 million Edge users, shipped version 5.5.0 on 9 July 2025 as a silent automatic update. That version captured conversations on eight AI platforms, ChatGPT and Claude among them, and sent them to the publisher for sale to advertisers.
All of these cases arrived through the store’s normal auto-update. A store badge, a big install count or an approved security product on your allowlist did not stop them, which is why the controls below watch versions and publishers as well as names.
Why don’t SWG, CASB, DLP or EDR catch an extension reading the page?
Each sits at a layer the extension’s work never crosses. The CSA’s April 2026 notes put it directly: these tools “typically lack native visibility into DOM-level operations that occur within a browser session,” and the extension “operates inside the trusted user session, with the trusted user’s access, over the trusted user’s network path” (CSA, 10 April 2026).
- A secure web gateway or CASB sees HTTPS leaving the user’s browser. A POST to a reputable-looking domain, sent from a signed-in employee’s machine, looks like any other web app. Security service edge inspection happens “at the TLS termination layer, before content is rendered in the browser” (CSA, 11 April 2026).
- Content-pattern DLP watches channels such as email, uploads and file copies. An extension reading the DOM moves nothing through those channels until it sends its own request, often encoded or batched, as in the 30-minute uploads from the fake sidebars.
- EDR sees the browser process. It can flag a known-bad extension ID or a malicious domain once someone has published them, but it can’t judge whether a legitimate AI helper reading a customer record is acceptable use. The CSA says EDR “without browser-specific instrumentation” has the same gap.
That leaves the browser’s own policy engine as the main control point. The other piece is visibility on the laptop for what users deliberately send to AI tools, which is where Jazz works on managed Windows and Mac fleets.
How do you stop data leaking through browser extensions employees install?
Run it in this order. Each step makes the next one smaller.
- Inventory every browser and every profile. Include Edge, Brave, Firefox, Arc and Safari, and personal Chrome profiles on work laptops. The open-source osquery
chrome_extensionstable reads Chrome, Edge, Brave, Arc, Vivaldi, Opera, Dia and Comet per profile, with each extension’s ID, version, author, install time and whether it came from the store. Itspermissionscolumn, though, doesn’t include Manifest V3host_permissions(osquery’s parser reads onlypermissionsandoptional_permissions), so query the hiddenmanifest_jsoncolumn to find all-sites access. Thefirefox_addonsandsafari_extensionstables list what is installed but have no permissions column, and the Safari table needs Full Disk Access. Sort the results by permission, not by name. Anything with all-sites access,cookies,clipboardRead,debuggerorwebRequestgoes to the top. - Manage by permission, and keep extensions off sensitive sites. Chrome’s
ExtensionSettingspolicy takesblocked_permissionsandruntime_blocked_hosts. Google’s admin article “Prevent Chrome extensions from altering webpages” (undated, accessed October 2026) says blocked hosts stop script injection, cookie access and request changes there, for up to 100 entries. Put your identity provider, email, CRM, HR, finance, admin consoles and AI chat domains on that list, withruntime_allowed_hostsexceptions for the few extensions that must work there. Microsoft’s “Detailed guide to the ExtensionSettings policy” for Edge (January 2024) documents the same keys. This is shorter to maintain than an allowlist of thousands of IDs and holds up when an allowed extension goes bad. Set it as machine policy through Group Policy or your MDM. Google’s “Set Chrome policies for users or browsers” (undated, accessed October 2026) says policy applied to a signed-in cloud user doesn’t follow a personal Gmail sign-in, while its “Understand Chrome policy management” (undated, accessed October 2026) says platform policies apply to every user of the device. To stop personal profiles appearing in the first place,BrowserAddPersonEnabledblocks adding new profiles andRestrictSigninToPatternlimits which Google accounts can become a profile’s primary account, per the Chromium policy definitions (accessed October 2026). - Block sideloading and developer mode. Chrome 128 added
ExtensionDeveloperModeSettings, and the Chromium policy definition (accessed October 2026) says a*blocklist also blocks all unpacked extensions. In Firefox, setinstallation_modetoblockedin the default entry and useinstall_sourcesto limit where installs come from. - Default-deny for high-risk groups. For admins, finance, HR and engineers with production access, block
*and allow named extensions by ID. These people hold the sessions and cookies worth stealing. - Run requests through a short review. Each approved extension gets a business owner, a reason, the permissions it was approved with and an expiry date for re-review. Prefer extensions that ask for
activeTab, which MDN describes as acting on the current page “when the user asks, without requesting broad host permissions.” - Watch for permission expansion and new owners. Chrome disables an extension until the user accepts new permissions, and adding host patterns triggers that prompt, per Google’s Chrome for Developers pages “Permission warning guidelines” and “Declare permissions” (both updated February 2024). A malicious update that keeps the old permissions triggers nothing, and ShadyPanda’s backdoor pulled new code every hour without any store update at all. Re-run the inventory on a schedule and diff version, permissions and author. The CSA recommends classifying extensions “by permission scope, publisher identity, update history, and store provenance.”
- Block or manage any browser you can’t put policy on. Brave’s Help Center (updated August 2026) says it “supports all of the policies that Chromium does.” Firefox takes the same ideas through ExtensionSettings, with
runtime_blocked_hostsandblocked_permissionsfrom Firefox 153 andrestricted_domainsbefore that. Safari extensions can be managed through Apple’s declarative configuration, only on supervised Macs running macOS 15 or later. Arc’s admin help page documents a per-user policy file and a few settings, none for extensions, and The Register reported in May 2025 that Arc was in maintenance mode, so treat it as a browser you can’t govern. Anything in that position, including portable browsers, needs application control or removal.
| Browser | Policy that controls extensions | Keep extensions off sensitive sites | Permissions-level inventory |
|---|---|---|---|
| Chrome | ExtensionSettings, ExtensionInstallBlocklist, ExtensionDeveloperModeSettings (Chrome 128+) |
runtime_blocked_hosts |
osquery chrome_extensions |
| Edge | Same keys (Microsoft guide, January 2024) | runtime_blocked_hosts |
osquery chrome_extensions |
| Brave | All Chromium policies (Brave, August 2026) | Chromium keys | osquery chrome_extensions |
| Firefox | ExtensionSettings |
restricted_domains; runtime_blocked_hosts from Firefox 153 |
Installed list only |
| Safari (macOS 15+, supervised) | Apple declarative Safari extensions settings | Per-extension allowed and denied domains | Installed list only |
| Arc | Not documented; Arc’s help pages show a per-user plist with policies such as URLBlocklist |
Not documented | osquery chrome_extensions |
For the bigger question of whether to move SaaS work into a separate managed browser, see do you need an enterprise browser.
Should you approve an AI sidebar extension, restrict it, or ban the category?
Approving one vetted assistant is usually the workable choice. People who want an AI sidebar and have no approved option go looking in the store, which is where the lookalikes are. The three positions look like this:
- Sanctioned single extension. Pick one assistant your company already has an enterprise contract with, force-install it, and block every other AI extension by ID or by permission. This only works if the enterprise terms cover the extension.
- On-click only. Allow an assistant that works through
activeTaband blocks its access to sensitive hosts withruntime_blocked_hosts, so it reads a page only when someone invokes it there. Check that claim yourself, because the UC Davis audit found some assistants ran automatically on searches. - Category ban. Fits regulated roles and the high-risk groups above, where a mistaken click on a patient portal or payroll page is too costly.
Before approving any of them, check four things in writing. Ask how long the vendor keeps what the extension sends, whether that data trains models, what the admin can control and where selected or visible text goes. The answers vary. Anthropic’s admin documentation for Claude in Chrome says zero data retention is “Not supported,” offers site allowlists and blocklists, and documents a forceLoginOrgUUID policy that limits the extension to your organization’s account (Anthropic). Its safety page says “whatever is visible in one of those tabs” becomes part of the conversation (Anthropic). Assistants that click and type for the user are a separate problem, covered in how to secure AI agents on employee computers.
What should you do when a bad extension is found on company laptops?
Assume it read everything its permissions allowed, for as long as it was installed.
- Remove it everywhere at once. Add the ID to the blocklist in every managed browser. Microsoft’s Edge policy documentation for
ExtensionInstallBlocklist(updated September 2026) says blocked extensions already installed “will be disabled”; Mozilla saysblockedin Firefox removes the extension from the device. Then re-query the inventory to confirm it is gone from every profile. - Scope the exposure. Use install time, version and profile from the inventory to list affected users and the window. In the Cyberhaven case only one version, live for 31 hours, was bad.
- Revoke sessions and rotate credentials. Stolen session cookies sign the attacker in without a password or MFA prompt. CISA’s eviction guidance says to “Revoke compromised tokens or update revocation list” and to force a password reset, while warning that reissuing tokens disrupts legitimate sessions. Include SaaS API tokens and AI API keys that appeared in pages or chats.
- Treat recent AI chats and visible pages as exposed. If the extension could read ChatGPT or the CRM, assume those conversations and records left. Pull the affected users’ recent AI history and the apps they had open.
- Decide on notification. If personal data was involved, GDPR Article 33 requires notifying the supervisory authority “without undue delay and, where feasible, not later than 72 hours” after becoming aware, unless the breach is unlikely to put people at risk. Check customer contracts and local breach laws too, and record the reasoning if you decide not to notify.
The broader runbook is in Jazz’s data leak incident response plan.
What does CIS Safeguard 9.4 require, and what will an auditor ask for?
CIS Controls v8 and v8.1 Safeguard 9.4 says: “Restrict, either through uninstalling or disabling, any unauthorized or unnecessary browser or email client plugins, extensions, and add-on applications.” It applies to Implementation Groups 2 and 3 under the Protect function (CIS Controls Assessment Specification).
A written policy doesn’t meet it. CIS’s own assessment specification measures 9.4 by collecting the installed plugins on each asset, comparing them with the authorized software inventory, and reporting “the percentage of assets compliant with authorized browser plugins.” That is the evidence to keep: the inventory export, the enforced policy and the compliance rate over time.
Is DLP delivered as a browser extension a safer answer?
It adds coverage in the browser it lives in, and it adds risk. A DLP extension needs the same page-wide access this article warns about, and the Cyberhaven incident showed that a security vendor’s extension can be hijacked through its publishing account like any other. It also covers only the browsers and profiles where it’s installed. A second browser, a personal profile or a private window it isn’t allowed into falls outside it, and MDN notes that extensions don’t run in private windows by default in Firefox. Put any DLP extension through the same review and allowlist as every other extension, and check which of your browsers and profiles it actually reaches. An endpoint agent is privileged software too and deserves the same vendor scrutiny. Jazz’s agent works at the operating-system level with no browser extension, so it adds nothing to the extension list and its reach doesn’t depend on which browser or profile someone picks.
Where Jazz fits, and what it doesn’t do
The agent captures copy and paste, GenAI prompts and file uploads with no browser extension to deploy, and Jazz tells a corporate account from a personal one on the same service. Melody looks at the data, where it came from and went, who moved it and whether it fits a business process, then delivers a pre-investigated answer. When action is needed Jazz responds in proportion, from a prompt to a targeted block.
The limits matter here. Jazz is not an extension inventory or blocking tool, and it doesn’t claim to see an extension silently scraping the DOM in the background; that stays with browser policy and the steps above. Phones, Linux and unmanaged devices are outside its reach.
Frequently asked questions
Are AI browser extensions that can read every page a data leak risk?
Yes. An AI extension with access to all sites can read email, CRM records, session cookies and AI chats in any open tab. A USENIX Security 2025 audit of nine GenAI assistants found they often send full page content to their own servers, and one captured a Social Security number typed into an IRS form.
How do I stop data leaking through browser extensions employees install?
Inventory extensions in every browser and profile, then use browser policy to block risky permissions and keep extensions off sensitive domains with runtime-blocked hosts. Block developer mode and sideloading, run a default-deny allowlist for admins, finance, HR and production engineers, and re-check versions, permissions and publishers on a schedule.
Why don't our web gateway, CASB, DLP or EDR catch a malicious extension?
The extension reads the page inside the user's signed-in session and sends data out as ordinary HTTPS. Gateways and CASBs see normal web traffic, content-pattern DLP watches channels such as email and uploads, and EDR sees the browser process. EDR can flag known-bad extension IDs but can't judge a legitimate AI helper reading a customer record.
What happened to the Cyberhaven security extension V3?
On Christmas Eve 2024 an attacker phished a Cyberhaven developer into approving a malicious OAuth app, then published version 24.10.4 of the extension. It was live for 31 hours on 25 and 26 December, could send session cookies to the attacker, and was one of about 20 extensions compromised by the same phishing campaign.
Should we approve an AI sidebar extension or ban them all?
Approving one vetted assistant usually works better than a ban, because people without an approved option go looking in the store, where the lookalikes are. Check data retention, training use, admin controls and where visible text goes. Restrict it to on-click use away from sensitive domains, and ban the category for high-risk roles.
What should we do when a malicious extension is found on company laptops?
Block and remove it in every managed browser, then confirm through inventory. Use install times and versions to scope affected users, revoke sessions and rotate passwords, API tokens and AI API keys. Treat recent AI chats and visible pages as exposed, and decide whether GDPR's 72-hour notification or other duties apply.
What does CIS Controls Safeguard 9.4 require for browser extensions?
Safeguard 9.4 in CIS Controls v8 and v8.1 requires restricting unauthorized or unnecessary browser and email client extensions, for Implementation Groups 2 and 3. CIS measures it as the percentage of assets with only authorized browser plugins, so keep the inventory export, the enforced policy and the compliance rate as evidence.
Is DLP delivered as a browser extension safe to deploy?
It adds another extension with page-wide access, and a DLP vendor's own extension was hijacked in December 2024. It also covers only the browsers and profiles where it is installed. Review it like any other extension. Jazz's endpoint agent needs no browser extension and works in any browser on managed Windows and Mac laptops.
Can Jazz detect or block malicious browser extensions?
No. Jazz is not an extension inventory or blocking tool, and it doesn't claim to see an extension silently scraping a page. On managed Windows and Mac laptops it sees what employees copy, paste or upload into AI tools in any browser, tells personal accounts from corporate ones, and Melody investigates each risky movement.
How do we see what employees paste into AI tools in browsers we can't put policy on, like Arc or a portable browser?
Browser policy reaches only browsers you manage, and Arc documents no extension policies. On managed Windows and Mac laptops, Jazz's OS-level agent sees copy and paste, GenAI prompts and file uploads in any browser or desktop app, so switching browsers doesn't route around it. It doesn't block extensions themselves; that stays with browser policy.
Is there DLP for AI tools that doesn't need a browser extension?
Yes. Jazz runs as one endpoint agent at the operating-system level on Windows and macOS, so it adds nothing to the extension list. It sees pastes and uploads into AI tools in any browser, tells a personal ChatGPT account from the corporate one, and Melody returns a pre-investigated incident. Linux and unmanaged devices need other controls.
Do browser extension controls work on phones, Linux or personal laptops?
Browser policy reaches only browsers you manage. Safari extension management needs supervised Macs on macOS 15 or later, and Arc documents no extension policies. Jazz covers managed Windows and Mac laptops only, so phones, Linux and unmanaged devices need other controls, and any browser you can't put policy on should be blocked or removed.



