Contents
An agentic AI browser is not a search engine with a chatbot bolted on: it is a program that sees what you see on screen, logs in with your saved sessions, and acts on its own — filling forms, opening email, making purchases — without you touching the mouse at every step. Tools like Perplexity Comet, Microsoft Copilot Mode in Edge or ChatGPT Atlas are already used inside businesses for repetitive tasks: comparing quotes, managing bookings, reviewing invoices. The problem, confirmed twice by independent security researchers in under three months, is that these browsers cannot tell an instruction from their user apart from an instruction hidden inside an ordinary webpage. And that gap has already been used to steal passwords without the user doing anything at all.
What an agentic AI browser actually is
It is not an extension that suggests text: it is an assistant with permission to operate the browser the way a person would, using already-open sessions (email, banking, CRM) to complete multi-step tasks with little or no supervision. The main ones in 2026 are:
- ChatGPT Atlas (OpenAI): launched in October 2025. OpenAI announced in July 2026 that it would stop working on August 9, 2026 — under ten months later — folding its functions into the ChatGPT desktop app.
- Perplexity Comet: still active, and along with Atlas the most studied by security researchers.
- Microsoft Copilot Mode (in Edge): built into a browser already used by default in many businesses.
- Claude for Chrome (Anthropic): in limited testing, with the same risks acknowledged by its own maker.
The attack that has already worked, twice in three months
Comet: a Reddit post steals a password at a glance
On August 20, 2025, Brave’s security team published a proof of concept against Comet. They hid instructions inside a “spoiler” tag on a Reddit post, invisible to anyone browsing normally. When a user asked Comet to “summarize this page,” the assistant also read those hidden instructions and executed them as if they were a legitimate command: it navigated to the Perplexity account settings, extracted the user’s email, triggered a one-time password email, opened Gmail, read that code, and posted both values back as a reply on the same Reddit thread, where the attacker could collect them. Brave explained that the underlying flaw is that, for an agentic browser, there is today no reliable way to separate what the user asked for from what any page it visits happens to say.
Atlas: memory that gets poisoned and follows you across every device
Two months later, on October 27, 2025, security firm LayerX (now part of Akamai) found a different flaw in Atlas: a CSRF-style attack that abuses the user’s already-open session to inject hidden instructions directly into ChatGPT’s persistent Memory, the feature meant to remember preferences across conversations. Once that memory is “tainted,” the malicious instructions survive across different sessions, different devices and even different browsers, because they travel with the account, not the machine. In the same research, LayerX tested Atlas against 103 real-world phishing attacks currently circulating: Atlas blocked only 5.8%, compared with 47% for Chrome and 53% for Edge — making it close to 90% more exposed than conventional browsers.
Even the most careful one wasn’t spared: Claude for Chrome
Even Anthropic, which has publicly stated that prompt injection is “the biggest risk facing browser-using AI tools” and that the chance of it happening “is not zero,” had its own flaw: outside researchers found that any subdomain matching Claude’s trusted-origin pattern could send commands directly to the Claude for Chrome extension, because of an overly permissive allowlist. The patch reduced the problem, but did not close it entirely.
Why this doesn’t get fixed with an update
Three different vendors, three different flaws, in under seven months. This isn’t one company missing a detail: it is that an agentic browser processes a user’s command and the content of any page it visits through the same channel, with no secure boundary between “trusted instruction” and “text that can say anything.” Until that architecture changes, every patch closes one specific case, not the category of attack. Waiting for “the next version to fix it” is not a security strategy — we’ve already seen this with AI agents in general, in the excess access they’re routinely granted and in the AI tools staff adopt without ever going through IT — and the agentic browser is simply the most direct entry point into a company’s real accounts.
What this means for a business
If an agentic browser connected to an employee’s account leaks customer data or corporate credentials, that is a security breach just as much as if a person had caused it — the underlying exposure doesn’t change because an AI assistant acted on the user’s behalf instead of a human clicking directly. And the risk isn’t limited to a single leaked data point: it’s every open session — email, CRM, banking, supplier portals — that browser can use without anyone in the company watching it happen in real time.
5 things to audit now, before the first incident
- Find out which AI browsers staff are actually using, not what the policy says is allowed.
- Check whether those browsers have open sessions in tools holding customer data: CRM, email, banking, supplier portals.
- Verify whether persistent memory is enabled (ChatGPT Memory or equivalents) and what company information is stored in it.
- Put in writing which tasks are never delegated to an autonomous agent: payments, sending personal data, access to contracts.
- Request an external audit before approving corporate use — not after the first incident.
Frequently asked questions
What is prompt injection in an AI browser? It’s a technique where an attacker hides instructions inside a webpage, a comment, or a document. The agentic browser reads them alongside the user’s real command and, unable to tell them apart, executes them as if they were legitimate.
Does this only affect Comet and Atlas, or any AI browser? It affects the entire category. Brave documented it in Comet, LayerX in Atlas, and even Anthropic’s own Claude for Chrome had its own flaw. Anthropic itself publicly acknowledges the risk never reaches zero.
Could an attack like this count as a data breach? If personal data belonging to customers or employees is exposed, yes — regardless of whether the action was carried out by an AI assistant rather than a person directly.
Should I ban these browsers at my company? Not necessarily, but you should decide deliberately: what tasks they’re allowed to perform, what sessions they can operate with, and what gets audited — instead of letting each employee decide on their own.
How do I know if these are already being used at my company without my knowledge? It’s the same problem as shadow AI: tools staff adopt because they save time, without going through any approval process. A security audit catches it before it becomes an incident.
Before someone else decides, decide for yourself
An agentic AI browser connected to your company’s real accounts isn’t just another productivity tool: it’s a new door into your data that no vendor has fully closed yet. At Optima Quantum we run security audits that review these tools too, before they cause an incident, as part of our cybersecurity services for businesses. If you want to know which AI browsers and agents are already operating inside your company, contact us.
Sources consulted: Brave, “Agentic Browser Security: Indirect Prompt Injection in Perplexity Comet” (Aug 20, 2025); LayerX / Akamai, “LayerX identifies vulnerability in new ChatGPT Atlas browser” (Oct 27, 2025); The Register, coverage of the Atlas memory vulnerability (Oct 27, 2025); The Hacker News and SecurityWeek, coverage of the “ClaudeBleed” flaw in Claude for Chrome; MacRumors, announcement of ChatGPT Atlas’s shutdown (Jul 10, 2026).
Get our new articles by email
One email per article. Unsubscribe anytime.