November 10, 2026

Agentic Browser Security Risks Start With Who's Really Clicking

No items found.

Key Takeaways

  • The biggest agentic browser security risk is identity: the AI agent acts inside an employee's logged-in sessions, so most audit trails can't say who clicked.
  • Prompt injection, session misuse, cross-tab exposure, back-end data flows, sidebar sprawl, and thin audit trails trace back to that blurred identity.
  • Whether an AI browser such as Comet is safe at work depends on which sessions its agent can reach and whether its actions are attributable.
  • Decide to allow, restrict, or govern per session and per application, then enforce it at the browser layer where the agent acts.

An agentic browser acts inside sessions your security stack already trusts

An employee asks the browser sidebar to file an expense report and email the summary, and a minute later both are done. Nothing looks unusual to the security team.

An agentic browser, AI sidebar, or agentic extension reads pages, plans steps, and acts on the user's behalf. The difference from a chatbot is where the work happens. In a December 22, 2025 post from OpenAI, the company described its Atlas agent taking actions "inside your browser, just as you would."

Upstream controls such as SSO, MFA, conditional access, and CASB evaluated a person before the session opened. Once it's open, the agent inherits that trust. Nearly every control in the stack checked the employee at the door — and most have no way to notice who walks through next.

Agentic browser security risks aren't tied to one product, so understanding how AI agents talk to your browser matters more than which brand is installed.

Six agentic browser security risks trace back to one blurred identity

Most security teams already know the attack names; the harder part is explaining why existing controls didn't flag them. For enterprises, the main agentic browser security risks are indirect prompt injection, session misuse, cross-tab exposure, back-end data flows, extension sprawl, and thin audit trails.

1. Indirect prompt injection turns web content into instructions

Hidden text in a page, email, or shared document can read to the agent like a user's command. The Center for AI Standards and Innovation (CAISI) at NIST describes agent hijacking as "a type of indirect prompt injection." The attacker doesn't need a password, just a paragraph the agent will read.

OpenAI said prompt injection "is unlikely to ever be fully 'solved,'" though it's optimistic about materially reducing the risk. When an injected instruction lands, the resulting action runs as the user.

2. The agent acts with the employee's session and permissions

Whatever the employee can approve, delete, or share, the agent can attempt. OWASP LLM06:2025 Excessive Agency traces this failure to excessive functionality, excessive permissions, and excessive autonomy. An agent tidying a shared drive can as easily change its sharing settings, and the log will typically show the employee did it.

3. Cross-tab access weakens the browser's oldest boundary

The same-origin policy has long kept one site's scripts away from another site's data. A June 30, 2026 University of Washington study found four of seven agentic browsers create ways to bypass it if a prompt injection succeeds. Only one was exploited end to end, yet an agent with HR and webmail in adjacent tabs can treat both as one context.

4. Page content and prompts flow to the AI back end

Summaries and actions usually send page content, which can include sensitive records, to the AI provider's cloud. Gartner analysts cited this cloud data flow among their reasons for caution, as CSO Online reported on December 8, 2025. For most teams, the hard part isn't knowing data leaves; it's knowing which data left and on whose instruction.

5. Extensions and sidebars multiply faster than inventories

Most of this capability arrives through unmonitored AI entry points, without a procurement ticket. A Gartner survey of 302 cybersecurity leaders, released November 19, 2025, found 69% of organizations suspect or have evidence employees use prohibited public GenAI. The figure covers GenAI broadly, not AI browsers specifically.

The Cloud Security Alliance noted in April 2026 an AI extension "operates inside the trusted user session, with the trusted user's access." It's one reason shadow AI is a visibility problem, not a blocking problem.

6. The audit trail records the user, not the agent

The Five Eyes cyber agencies published joint guidance on agentic AI on May 1, 2026. It warns actions executed under a trusted identity produce "audit logs that appear legitimate and delay detection." It isn't browser-specific.

Proxies, EDR, and CASB were built for network traffic, endpoint processes, and app sessions, and they served their era well. Telling a human click from an agent's click inside an encrypted session wasn't a design criterion then. In our view, this is the core gap: most enterprises can't reliably tell user actions from agent actions.

Is it safe to use Comet or Atlas at work? Only where you can tell the agent from the employee

Employees are asking for these tools by name, and security teams want an answer they can defend in front of the board. The honest answer is yes, but only for low-risk work where the agent can't reach sensitive sessions and its actions are logged separately. Until both conditions hold, treat any AI browser as unsafe for regulated data.

OpenAI shut down ChatGPT Atlas on August 9, 2026, and 9to5Mac reported its browser tools moved into the new ChatGPT desktop app. The capability didn't leave; it changed address. A policy written against a product name ages in months, while one written against session access holds up.

In December 2025, Gartner analysts Dennis Xu, Evgeny Mirolyubov, and John Watts advised organizations to "block all AI browsers for the foreseeable future." CSO Online reported they still allowed small pilots on low-risk use cases that are easy to verify and roll back. That advice reflected the late-2025 market, so it reads best as a dated snapshot rather than a standing verdict.

Comet has moved on since then as well. Perplexity says Comet Enterprise, announced March 17, 2026, lets admins limit the agent to answering questions or allow actions only on approved domains. Those controls and the audit logs Perplexity describes are useful, though they're configured inside one browser rather than across the browser estate.

The joint Five Eyes guidance points the same direction, recommending agentic AI only for "low-risk and non-sensitive tasks." Blocking buys time, but it doesn't buy visibility. Most organizations will end up governing these tools, and the open question is whether they get to choose where.

A defensible late-2026 posture approves a short list of agentic tools for low-risk sessions and blocks the rest until attribution is in place. It keeps the yes small enough to watch while the controls catch up.

Allow, restrict, or govern based on the session the agent can reach

"Allow or block" is where most acceptable-use policies start, and it's the wrong unit of decision for agents. The better unit is the session, meaning what the agent can see and do inside each logged-in application.

Public research and summaries of non-sensitive pages can usually be allowed. Internal SaaS fits a restricted tier, typically read-only with no form submission. Finance, HR, admin consoles, and anything holding regulated data call for tight governance or a block.

In the governed tier, the agent can read a pay stub in the payroll application to explain a deduction to the employee who asked. Any change to bank details, salary, or tax withholding waits for that employee to confirm it in their own session. Accounts payable follows the same pattern, with the agent matching invoices to purchase orders and flagging mismatches while payment release stays a human click.

A sales rep might let the agent summarize a prospect's public filings, read the CRM record, and draft a follow-up. The send button, and any change to the record, stays with the human. Before you assign a tier, run each application through six questions:

  1. Reach: Which logged-in applications can the agent see or act in from this browser profile?
  2. Action scope: Can you enforce read-only or domain-limited actions for each application?
  3. Attribution: Will your logs show whether the employee or the agent performed each action?
  4. Data path: What page content leaves for the AI provider, and can you mask or block it?
  5. Credentials: Can the agent use saved passwords or tokens, or are credentials kept out of its context?
  6. Reversibility: Can your team detect and roll back a mistaken agent action quickly, as Gartner's pilot criteria asked?

The answers translate directly into an acceptable-use policy: which applications sit in which tier, which actions need human confirmation, and who reviews agent activity.

The fastest readiness test fits in an afternoon. Pick one agent-assisted action from yesterday and ask your SOC to prove whether a human or the agent performed it. If they can't, the answer for sensitive apps is "not yet," whichever AI browser is involved.

The riskiest deployments are often the sanctioned ones, which surprises many teams. An approved sidebar inside the corporate browser has more trusted sessions within reach than a stray AI browser on a personal profile. That's why governing what agents do, not just what they're told, has to follow the session rather than the install.

Browser-layer controls separate the agent from the employee it works for

Once you decide per session, you need a control point that can see each session and each actor inside it. An enterprise browser with built-in security sees the page, the session, and who's acting, which gives it the most complete vantage point.

Extensions add those controls to browsers in use, including AI browsers, and work well as a complement where a full browser rollout doesn't fit. With extensions alone, coverage depends on the extension being present and current in each browser, which is easier to promise than to prove. Network and proxy controls still earn their place, though they were designed for an era when the network was the perimeter.

Proxies see traffic, not who clicked inside an encrypted session. Whichever mix you choose, these are the browser-layer controls worth asking for:

  • Identity binding that ties each agent action to the user, group, device posture, and network.
  • User-versus-agent attribution recorded on every entry in the audit log.
  • Per-agent, per-application policy, including a read-only mode.
  • Credential isolation, so passwords stay out of the model's context.
  • Last-mile controls on clipboard, uploads, downloads, print, and screen capture.
  • Inline detection of prompt injection and data leakage.

Sequencing matters as much as the control list. A practical first step is turning on user-versus-agent logging for the sidebar you've already approved, without changing its permissions. A few weeks of those logs show which applications agents touch, which actions they attempt, and where employees step in.

With that baseline, you can widen read-only access where the activity looks clean and add confirmation steps where it doesn't. Expanding agent permissions before attribution is in place means granting trust you can't yet audit.

These controls are the design center of the Island Enterprise Browser. Island's Enterprise Browser for Agents tells the agent apart from the user and applies different policy to each. Agent work lands in the same audit trail as human work, so your SOC can pass the afternoon test with a log line.

Island Enterprise AI, which includes the Enterprise AI Browser, extends runtime protection to agents working beyond the browser. The Island Extension brings the same policy to Chromium-based AI browsers such as Comet, keeping security built in, not bolted on, wherever the agent runs.

Your audit trail should be able to name who clicked

Request a demo to pressure-test your agentic browser policy against your own sessions, and we'll walk through what we've built. The easiest agent to say yes to is the one your logs can name.

FAQs

What are the security risks of agentic browsers for enterprises?

Agentic browsers let an AI agent read untrusted content and act inside employees' logged-in sessions. That combination enables prompt injection, data exposure, and actions the audit trail attributes to the person.

Is it safe to use AI browsers like Comet or Atlas at work?

Atlas stopped working on August 9, 2026, so the question now applies to Comet and the browser tools in the ChatGPT desktop app. The safe zone is low-risk tasks where the agent can't reach sensitive sessions and its actions are logged separately.

Should enterprises block AI browsers, as Gartner recommended?

Blocking is a reasonable stopgap, and Gartner's December 2025 advice still allowed narrow pilots. Over the long term, session-level governance holds up better than blocklists built on product names.

Can prompt injection in AI browsers be fully fixed?

OpenAI calls it unlikely to ever be fully solved, and the UK NCSC warned in December 2025 it "may never be totally mitigated." The practical goal is limiting what an injected agent can reach and making its actions attributable.

Island Team

Island is defining the future of work for people and AI agents. Its enterprise agentic control plane helps organizations enable, govern, and audit agentic workforces alongside people. Island boosts productivity across devices, browsers, applications, networks, and data while protecting sensitive information, simplifying access, and helping enterprises scale AI safely.