AI Browsers Compared: Perplexity Comet, ChatGPT Atlas, Dia, Gemini in Chrome and Edge Copilot, and the Prompt-Injection Risk

A guide to AI-first and AI-enhanced browsers: what Perplexity Comet, OpenAI's ChatGPT Atlas, The Browser Company's Dia, Gemini in Chrome and Microsoft Edge's Copilot features do, what agentic browsing is good for, why indirect prompt injection is a serious risk, publisher concerns and safe-use advice.

PUBLISHED // UPDATED
DURATION // 06 MIN READ
AI Browsers Compared: Perplexity Comet, ChatGPT Atlas, Dia, Gemini in Chrome and Edge Copilot, and the Prompt-Injection Risk
FIG.01 // VISUAL DOCUMENTATION AI Browsers Compared: Perplexity Comet, ChatGPT Atlas, Dia, Gemini in Chrome and Edge Copilot, and the Prompt-Injection Risk

For two decades the web browser barely changed: type an address or a search, get a list of links, click and read. Since 2025, that has started to shift. AI companies have launched browsers built around assistants that summarise pages, answer questions across tabs and, in “agent” modes, click through websites to complete tasks. The incumbents have responded by adding assistants to Chrome and Edge. This guide compares the main options, explains what agentic browsing is good for and sets out the security risks that make caution sensible.

The main AI browsers

BrowserDeveloperLaunchedNotable features
CometPerplexityJuly 2025 (Windows, macOS); free download from October 2025Assistant beside every page; multi-step tasks; built on Chromium
ChatGPT AtlasOpenAIOctober 21, 2025 (macOS)ChatGPT built into browsing; optional agent mode on paid plans
DiaThe Browser Company (now part of Atlassian)2025AI-first successor to Arc, built around chatting with tabs
Gemini in ChromeGoogleRolling out from 2025Assistant answers questions across tabs; AI Mode in the address bar
Copilot in EdgeMicrosoftOngoingCopilot sidebar and Copilot Mode for page-aware help and tasks

Sources: Wikipedia’s articles on Comet and ChatGPT Atlas, Atlassian’s announcement and Google’s Chrome announcement.

Perplexity Comet grew out of Perplexity’s answer engine, which replaced lists of links with synthesised answers and citations. Comet puts that assistant next to every page and can carry out multi-step tasks within the browser. It launched on Windows and macOS in July 2025 for top-tier subscribers, became a free download in October 2025, and later arrived on Android and iOS.

ChatGPT Atlas brings ChatGPT into the browser. OpenAI released it for macOS on October 21, 2025, with versions for other platforms promised later. It uses a freemium model, with capabilities such as agent mode, which can carry out tasks on websites, reserved for paid plans.

Dia comes from The Browser Company, maker of the design-led Arc browser. The company stopped adding features to Arc to focus on Dia, an AI-first browser built around asking questions of your open tabs, and Atlassian announced in September 2025 that it would acquire the company.

Gemini in Chrome is Google’s answer: an assistant that can answer questions and find information across multiple tabs, plus an AI Mode in the address bar for complex questions. With Chrome’s huge installed base, this may reach more people than any standalone AI browser.

Microsoft Edge integrates Copilot into a sidebar and an optional Copilot Mode that can understand the page you are viewing and help with tasks such as comparing products.

What AI browsing is good for

Summarising and research. Asking for the key points of a long article, comparing information across several open tabs or getting an overview of a topic with citations are the most reliable uses.

Reading assistance. Explaining jargon, translating pages and answering questions about a document you are viewing.

Simple, low-stakes tasks. Filling in a form with information you provide, gathering product details or drafting a reply, with you reviewing before anything is submitted.

Not yet suitable without close supervision: tasks involving payments, banking, email accounts, account settings or anything hard to undo. Agents can misread pages, click the wrong thing or be manipulated, as described below.

The security problem nobody has solved

Agentic browsers read untrusted web content and can act with the user’s logged-in sessions. That makes them vulnerable to indirect prompt injection: hidden instructions on a page can steer the agent. Brave’s security team published a demonstration of this class of attack against Comet in August 2025 (Brave). OWASP lists prompt injection as the top risk for LLM applications (OWASP GenAI).

How the attack works

An AI browser agent combines three things: it reads content from any website, it can take actions such as clicking, typing and navigating, and it often does so while you are logged in to your accounts. A malicious page, comment, email or even an image can contain text written as instructions to the AI, sometimes hidden from human readers with white-on-white text or tiny fonts. If the agent treats that text as a command rather than as data, it might open another tab, read information from a logged-in site and send it elsewhere, or take an action you never asked for. Unlike traditional web attacks, this does not require breaking the browser’s security model; it exploits the model’s difficulty in separating trusted instructions from untrusted content.

Vendors have added mitigations, such as asking for confirmation before sensitive actions, restricting agents on certain sites, logging out of accounts during agent tasks and training models to resist injected instructions. Researchers continue to find bypasses, so these reduce the risk rather than remove it.

A practical defence is separation. Use agent features in a separate browser profile that is not logged in to sensitive accounts, keep the agent’s permissions narrow and review every action that sends messages, makes purchases or changes settings.

What is at stake for search and publishers

AI browsers and answer engines change who controls the gateway to the web. Search advertising depends on people clicking results; assistants that answer questions directly reduce clicks to the sources they draw on. Publishers worry that this deprives them of traffic and revenue, and several have sued AI companies or signed licensing deals. How answer engines attribute and compensate sources will shape the economics of web publishing, and it is one reason the incumbents, whose businesses rely on search traffic, are adding AI carefully rather than abandoning links.

Choosing an AI browser

If youConsider
Want AI features without switching browsersGemini in Chrome or Copilot in Edge
Already rely on Perplexity for researchComet
Use ChatGPT heavily and are on macOSChatGPT Atlas
Liked Arc’s design and want an AI-first workflowDia
Handle sensitive accounts and dataStay with a conventional browser profile for those, and use AI features separately

Our analysis

Traditional browsers are not dying. Chrome, Safari and Edge are adding assistant features of their own, and distribution usually beats novelty. The realistic near-term change is narrower: AI browsers are genuinely useful for summarising, comparing and research. Letting an agent act on banking, email or shopping accounts should wait until vendors can show robust isolation between page content and user commands. Until then, use agent features in a separate profile that is not logged in to sensitive accounts, and review every consequential action yourself.