In October 2025, OpenAI launched Atlas, whilst Perplexity made Comet free to use, removing the $200 monthly subscription fee. Google is now also offering its Chrome browser augmented with Gemini AI.
Unlike traditional browsers, these integrate AI capabilities directly into the browsing environment. They can summarise content, answer questions in context and, in some cases, act semi-autonomously to perform tasks such as research or shopping for the user.
ChatGPT Atlas
OpenAI’s Atlas extends ChatGPT into a browser, although it is currently limited to modern Mac users (those running macOS with Apple Silicon) and not universally available. It leverages existing ChatGPT technology to analyse or summarise webpages or extract key data, allowing users to ask a question and interact with the chatbot without switching between windows. In the preview Agent Mode, the browser is also able to take limited autonomous actions on the user’s behalf.
OpenAI insists that browsing data is not used for AI training unless users opt in, but questions remain about what contextual information the system retains during a session. Atlas has also faced scrutiny following a reported security vulnerability (CVE) involving session persistence.
Comet
Perplexity’s Comet browser is now generally available and free to use. Unlike Atlas, it was designed as an AI-first browser rather than a conventional one with added tools. Built on Chromium, it integrates Perplexity’s search and reasoning engine throughout. Users can pose questions about any open page, compile summaries across tabs and automate repetitive research.
However, Comet’s deep integration also magnifies attack surfaces. Because the model interprets text directly from webpages, there is a risk of prompt injection i.e. where malicious actors embed hidden instructions to manipulate the model’s behaviour, exfiltrate data or trigger unwanted actions. Combined with uncertainties about how Perplexity stores or filters browsing data, these issues highlight the overlap between AI safety and cybersecurity.
Chrome with Gemini
Instead of creating a new product, Google has embedded Gemini directly into Chrome on macOS and Windows. Users can now activate an ‘AI mode’ in the address bar to generate summaries, explanations or navigation prompts. Google emphasises that this operates within its existing security framework, including Enhanced Safe Browsing and automatic credential protection.
Yet underlying risks remain: Chrome’s scale ensures rapid adoption, but it could also multiply exposure if vulnerabilities do emerge.
Overall, the distinction between searching with AI and asking AI to search seems uncertain. Embedding a model within the browser may assist professionals and specialists (e.g. buyers, analysts, researchers) handling large volumes of information or very specific tasks, but general users may find the gain in speed or clarity to be marginal.
The emergence of agentic capabilities in AI browsers is a more interesting technical prospect: Agent Modes promise to string actions together, navigate interfaces, and make decisions under broad user intent. This could similarly prove useful for specific, repetitive tasks, but currently seems less useful for general browsing. They also expose the brittleness of current models. Minor ambiguities in a prompt or webpage can cause the agent to pursue the wrong objective, loop indefinitely, or trigger actions that exceed what the user intended. For now, agentic behaviour is an interesting prospect but currently far from reliable.
The most significant impact may lie in data and security. While mitigations exist, the risk is inherent to any AI operating at this level of access. AI browsers mediate both what users read and what the models see, merging personal browsing data with cloud inference. Yet none of these providers have fully demonstrated the necessary transparency, safeguards and opt-outs for us to underestimate the importance of both security and privacy risks.