AI browser security risks: what AI browser agents and AI extensions mean for your company
The main security risk of AI browser agents is that they act inside the user's authenticated sessions: whatever the employee can do in email, the CRM or the bank portal, the agent can do too, and a web page can try to give it instructions through prompt injection. The vendors themselves say prompt injection cannot be fully solved. AI extensions add a second risk, because many of them can read every page the user opens, and some have been caught sending AI conversations to third parties.
You can reduce both risks with controls you already have: enterprise policies for the agents built into Chrome and Edge, admin settings for third-party agents, an allowlist for extensions and data loss prevention for the most sensitive data.
This guide sets out the risks with dated, sourced examples, the permissions that matter, and the controls that exist as of October 2026.
What is an AI browser agent?
An AI browser agent is software that reads web pages and takes actions in the browser for the user: clicking, typing, filling forms, moving between tabs. It comes in three forms:
| Form | Examples (status October 2026) |
|---|---|
| Built into the browser | Gemini in Chrome with auto browse (Google); Copilot in Edge, with "Browsing with Copilot" for business in preview since Edge 148 (Microsoft) |
| Extension in an existing browser | Claude in Chrome (Anthropic), generally available since 26 August 2026 on paid Claude plans, Chrome only |
| Standalone AI browser | Perplexity Comet (Mac, Windows, iOS, Android, with Comet Enterprise); Opera Neon; Dia (The Browser Company, acquired by Atlassian) |
OpenAI's ChatGPT Atlas, launched for macOS in October 2025, was one of the first examples and much of the 2025 research was done on it. OpenAI has since retired it: its help center says Atlas no longer receives updates, and its browsing features moved to a built-in browser in the ChatGPT desktop app.
Gartner defines an AI browser as an AI sidebar plus the ability to carry out transactions on its own, according to The Register's coverage of its advice (see below).
Why agents are different: they act as the user
Traditional web security keeps one site from reading another. An agent sidesteps that, because it runs inside the user's own browser profile, with the same cookies, IP address and device. Brave's security team put it bluntly in August 2025: the AI operates with the user's full privileges across authenticated sessions, which makes protections like the same-origin policy and CORS effectively useless against this kind of attack (Brave).
That has a consequence for your logs. Zenity Labs wrote in December 2025 that, from a website's point of view, agent-initiated actions are indistinguishable from those of the authenticated user (Zenity Labs). WorkOS made the same point in March 2026: from the server's perspective, it is just you clicking around. Some cloud-hosted agents can identify themselves by signing their requests, for example through Cloudflare's signed agents programme (Cloudflare, August 2025), but that covers remote browsers, not an agent running in the employee's own Chrome. In practice, your SaaS audit trail will attribute the agent's actions to the employee.
Prompt injection: how a web page hijacks the agent
Prompt injection is when content the model reads changes what it does. OWASP lists it first in its Top 10 for LLM applications and separates direct injection (the user's input) from indirect injection (instructions hidden in a website, file or email the model processes). Agents are exposed to the indirect kind on every page they read.
Simon Willison's "lethal trifecta" (June 2025) explains why browsers are the worst case: an agent becomes dangerous when it has access to private data, exposure to untrusted content and a way to communicate externally. A browser agent signed in to work accounts has all three by design.
The demonstrations so far, in date order:
- Comet and a Reddit comment (Brave, 20 August 2025). Instructions hidden behind a spoiler tag in a Reddit comment. The user asked Comet to summarise the page; the agent read the user's email address from their Perplexity account, requested a one-time password, opened Gmail where the user was signed in, read the code and posted both in a reply to the comment. Account takeover, with no further user action.
- Scamlexity (Guardio Labs, 20 August 2025). Asked to buy an Apple Watch, Comet completed a purchase on a fake Walmart site, autofilling the saved address and credit card in some runs. In another test it followed a phishing email and helped fill in bank credentials on a live phishing page (Guardio).
- CometJacking (LayerX, 4 October 2025). A crafted URL carried a prompt that made Comet pull data from connected services such as Gmail and Calendar, encode it in base64 to slip past checks and send it to an attacker's server.
- Unseeable prompt injections (Brave, 21 October 2025). Faint text in an image, invisible to the user, was passed to the model when the user took a screenshot and asked about it.
- Atlas phishing test (LayerX, 27 October 2025). Against 103 real phishing and malicious pages, Atlas stopped 6 (5.8%), compared with 53% for Edge and 47% for Chrome. LayerX also showed a cross-site request forgery that planted hidden instructions in ChatGPT's memory.
- Comet audit (Trail of Bits, 20 February 2026). In a pre-launch audit for Perplexity, four injection techniques exfiltrated emails from Gmail (Trail of Bits).
A corporate example is easy to imagine without any of these: an AI agent asked to screen job applications reads a CV with hidden text telling it to rank the candidate first. The mechanism is the same.
The vendors say it cannot be fully solved
This is not a criticism from outside. The people building these products say it themselves:
- OpenAI, December 2025: prompt injection, like scams and social engineering, is unlikely to ever be fully "solved" (as quoted by TechCrunch).
- UK NCSC, December 2025: prompt injection may never be totally mitigated, because a language model does not separate data from instructions (NCSC blog).
- Google, in its help page for Gemini auto browse: safeguards do not guarantee protection against prompt injection (Google).
- Perplexity calls it a frontier security problem in its post on mitigating prompt injection in Comet (October 2025).
Anthropic is the only vendor that publishes attack success rates for its browser agent. In August 2025 it measured 23.6% attack success in autonomous mode without mitigations and 11.2% with them. On a harder test published in August 2026, the rate fell from 17.6% (Claude Opus 4.5, no extra safeguards) to 0% for Claude Opus 5 with probes and a safety classifier (Anthropic). Anthropic says the two tests are not comparable. The direction is encouraging; zero on a benchmark is still not a guarantee on the open web.
AI extensions: the quieter risk
Agents get the headlines, but ordinary AI extensions are already a proven data leak. You do not need an agent to read a ChatGPT conversation; any extension that can read the page can.
| Date | Incident | Scale |
|---|---|---|
| March 2023 | "Quick access to Chat GPT", a working ChatGPT client that stole session cookies and took over Facebook business accounts (Guardio Labs) | Over 2,000 new installs a day |
| December 2024 | Cyberhaven's extension compromised through a phished developer account; the malicious version targeted logins to social media advertising and AI platforms | About 400,000 corporate users |
| July 2025 (published December 2025) | Urban VPN Proxy began harvesting prompts and responses from ChatGPT, Claude, Gemini, Copilot, Perplexity and others, on by default (Koi) | 6 million Chrome users for Urban VPN alone, with a Featured badge |
| December 2025 | Two fake AI sidebars sent full ChatGPT and DeepSeek conversations and all tab URLs to a server every 30 minutes (OX Security) | Over 900,000 combined |
| August 2026 | One of those sidebars returned to the Chrome Web Store and shipped a new payload in July 2026 (Netskope) | 200,000 users |
LayerX also showed in July 2025 that an extension with access to an AI tool's page, even without special permissions, can read and write prompts in ChatGPT, Gemini, Copilot, Claude and DeepSeek. That was a proof of concept, not an attack seen in the wild, but it shows that "it is just a writing helper" is not a security assessment. For how to find what is installed, see our guide to shadow AI browser extensions.
Which permissions should worry you?
These permissions decide what an AI extension or agent can actually do. The warnings are the texts Chrome shows at install, from the permissions list and Chromium's source.
| Permission | Why it matters | Chrome's warning |
|---|---|---|
Host access to all sites (<all_urls>) |
Read and change every page, combined with scripting and cookies | "Read and change all your data on all websites" |
debugger |
Drive the browser: click, type, take screenshots | "Access the page debugger backend" plus the all-sites warning |
scripting |
Inject code into pages (needs host access) | None of its own |
cookies |
Read session cookies on allowed sites | None of its own |
nativeMessaging |
Talk to applications installed on the computer | "Communicate with cooperating native applications" |
tabs |
See the URLs and titles of open tabs | "Read your browsing history" |
Two details are worth knowing. When an extension attaches the debugger, Chrome normally shows a bar saying the extension "started debugging this browser", but according to Chromium's source the bar is suppressed for extensions installed by enterprise policy. And a legitimate agent looks like a risky extension on paper: Anthropic's own documentation lists debugger, scripting, tabs and nativeMessaging among the 15 permissions Claude in Chrome uses. Permissions tell you what an extension could do, not whether you trust its publisher, so you need both checks.
What Gartner advises
According to The Register's report (8 December 2025) on Gartner's research note "Cybersecurity Must Block AI Browsers for Now", Gartner advised CISOs to block AI browsers for the foreseeable future. As reported, the analysts warned that default AI browser settings prioritise user experience over security, that sensitive data is often sent to the cloud AI back end, and that agents could be deceived into navigating to phishing sites. For organisations that do allow them, the reported advice was to assess the back-end AI, train users not to leave highly sensitive data open alongside the sidebar, restrict the agent's email capability and define prohibited use cases.
You do not have to agree with a blanket ban to use the checklist. It is a reasonable starting point for a pilot.
How to control the agents built into Chrome and Edge
Both browsers have specific policies. Set them explicitly rather than relying on defaults.
Chrome (Gemini in Chrome), per Google's admin guidance:
GeminiSettings: master switch for Gemini in Chrome (1 disables it).GeminiActOnWebSettings(Chrome 143+): controls auto browse, the agentic part; 1 blocks it. The policy text treats unset as allowed while Google's admin page says auto browse is off by default for managed users, so set it.GeminiActOnWebAllowedForURLsandGeminiActOnWebBlockedForURLs(Chrome 146+): per-site lists. If a URL matches both, Chrome allows it.GeminiSparkSettings(Chrome 150+): whether the Gemini Spark agent may use auto browse in Chrome. The policy description itself warns of significant security risks.
Edge (Copilot), per Microsoft Learn:
AllowBrowsingWithCopilot(Edge 148+): enables or disables agentic browsing; it requires a Microsoft 365 Copilot subscription.BrowsingWithCopilotAllowListandBrowsingWithCopilotBlockList: if the allow list is empty, the feature is unavailable on all sites. In Edge the block list wins over the allow list, the opposite of Chrome.CopilotCoworkToolActionsEnabled(Edge 152+): whether Copilot Cowork can take actions in Edge.
There is no Edge policy called "CopilotModeEnabled", despite what some blogs say. Microsoft also says agentic browsing with Copilot for business is available worldwide with a Microsoft 365 Copilot licence, excluding the European Economic Area (Edge blog, May 2026).
Controls for third-party agents and AI browsers
- Claude in Chrome. On Team and Enterprise plans, owners control it under Organization settings > Claude in Chrome, with an "Enable for your team" toggle that is on by default, plus a site allowlist and blocklist. The Chrome policy
forceLoginOrgUUIDpins the extension to your organisation (Anthropic admin controls). Anthropic recommends a restrictive allowlist and a limited pilot. - Perplexity Comet. Comet for Enterprise offers agent permission controls, including website restrictions and action approvals, and supports Chromium policies under its own macOS bundle ID,
ai.perplexity.comet(Perplexity help). - Standalone browsers in general. Windows App Control for Business or AppLocker can block unapproved executables, and the Edge management service has a "Block other browsers" option that requires Intune.
- Extensions. An allowlist with
ExtensionInstallBlocklistset to*, plusExtensionSettingsto block thedebuggerpermission where it is not needed and keep extensions off sensitive internal sites withruntime_blocked_hosts. From Chrome 154,runtime_blocked_hostson an extension withdebuggerdisables debugger attach entirely. The full walkthrough is in our guide to blocking AI extensions with Chrome and Edge policies.
Data loss prevention for AI
DLP limits the damage when something goes wrong. Microsoft Purview browser DLP in Edge for Business can audit or block pasting, uploading, copying and printing to consumer AI apps such as ChatGPT, Gemini, DeepSeek and Perplexity, on Intune-managed Windows devices, with pay-as-you-go billing (Microsoft Learn). Claude is not in its supported list as of October 2026. Google's Chrome Enterprise Premium can use URL filtering to block unapproved AI tools and apply upload and copy/paste restrictions, and in August 2026 Google said its DLP rules are enforced on agentic actions as on user actions (Google Cloud blog).
A practical checklist
- Inventory the AI extensions and AI browsers in use, and note their permissions.
- Decide which agents you allow, for which teams and which sites. Keep agents off banking, HR and admin consoles to start with.
- Set the built-in agent policies in Chrome and Edge explicitly.
- Enforce an extension allowlist, and block
debuggerand broad site access where approved tools do not need them. - Use vendor admin controls for any agent you approve: site lists, org pinning, approval modes.
- Add DLP for the data you cannot afford to leak.
- Train users. Agents should get specific instructions, not "do whatever is needed", and approval prompts deserve attention, not reflex clicks.
- Record approved tools in your AI register, as described in our guide to the ISO 42001 AI inventory.
For step 1 on individual machines, we built a free tool. In the interest of transparency: Wake is made by Osabee, the publisher of this guide. It is a Chrome and Edge extension that lists the installed extensions, separates AI agents (AI tools that can act on pages or read every site) from AI assistants and other high-privilege extensions, and shows the reasons: debugger control, code injection, session cookies, native messaging, installs from outside the store. It judges by declared permissions rather than behaviour, sees one profile at a time and has no central dashboard yet. Get it on the Chrome Web Store or Edge Add-ons.
Frequently asked questions
Is it safe to use AI browser agents at work?
AI browser agents are not risk-free at work, because they act with the employee's logged-in access and can be manipulated by prompt injection hidden in web pages. OpenAI, the UK NCSC and Google have all said prompt injection cannot be fully prevented. Agents can be used more safely on a short list of approved sites, with approval prompts on, enterprise policies set and sensitive systems excluded.
Is prompt injection still possible in 2026?
Prompt injection is still possible in 2026. Vendors have added classifiers, confirmation prompts and separate models that check each action, and Anthropic reported 0% attack success for Claude Opus 5 with its safeguards on a test published in August 2026. But OpenAI and the UK NCSC have both said prompt injection may never be fully solved, so companies should treat it as a residual risk.
Is the Perplexity Comet browser safe for business?
The Perplexity Comet browser has been the subject of several published attacks, including Brave's Reddit comment takeover in August 2025, Guardio's fake shop purchase and LayerX's CometJacking in October 2025, and a Trail of Bits audit in February 2026. Perplexity has added classifiers and confirmations, and Comet Enterprise offers admin controls. Whether it is acceptable depends on your sites, data and policies.
Are browser extensions a security risk?
Browser extensions are a security risk when they have broad permissions, such as access to all websites, cookies or the debugger, because they can read everything the user sees in the browser. In December 2025, researchers found extensions with millions of users collecting conversations with AI chatbots. An allowlist of approved extensions and a review of permissions reduce the risk considerably.
What are the biggest security risks of AI agents in the browser?
The biggest security risks of AI agents in the browser are indirect prompt injection from web pages, actions taken with the user's full logged-in privileges, data exfiltration to attacker-controlled sites, and phishing pages the agent fails to recognise. A further issue is accountability: websites and audit logs usually cannot tell an agent's actions apart from the user's own clicks.
What is an agentic browser?
An agentic browser is a web browser with a built-in AI agent that can act on pages for the user, for example filling forms, comparing products or sending messages, instead of only answering questions. Examples include Perplexity Comet, Opera Neon and Dia, while Chrome and Edge add agent features through Gemini and Copilot, and Claude in Chrome adds one as an extension.
Should companies block AI browsers?
Gartner advised CISOs in December 2025 to block AI browsers for the foreseeable future, according to The Register's report. One middle path is to block standalone AI browsers by default, set the Chrome and Edge agent policies explicitly and pilot one approved agent on a limited list of sites, with data loss prevention and an extension allowlist in place.
How can I protect my company from prompt injection in browser agents?
To protect a company from prompt injection in browser agents, limit what the agent can reach and do. Restrict agents to an allowlist of sites, keep approval prompts on for sensitive actions, give specific instructions, block agents on banking, HR and admin systems, and apply data loss prevention. The UK NCSC also recommends limiting tool power and logging inputs, outputs and tool calls.