ChatGPT Atlas & Perplexity Comet. Are AI Browsers Worth It? – AI News – #4 October 2025

For years, our window to the internet world remained unchanged. Whether we use Chrome, Firefox, Safari, or Edge, the idea is the same: an address bar, bookmarks, and a search engine that serves us a list of links. Artificial intelligence was at best an add-on, a plugin, or a side panel. That era is now coming to an end.

Two powerful players are entering the market: Perplexity Comet and ChatGPT Atlas from OpenAI. These are not ordinary browsers integrated with a chatbot. These are platforms where AI is the core, and traditional browsing becomes a secondary function. Is this an exciting novelty and a natural evolution of LLMs, or a real threat to our habits and security?

Perplexity AI has made a move that could shake up the market. Their flagship browser Comet, previously available only as an elite subscription for $200 a month, has now been made completely free.

The decision to remove fees is a strategic maneuver. Perplexity hopes to rapidly build a user base and engage in real competition with giants like Google and OpenAI. Free access is intended to accelerate adoption and gather invaluable data about user interactions with an AI-powered browser.

Comet is not an “AI browser.” It is “AI in the form of a browser.” The assistant is integrated into every new tab, proactively supporting users in everyday tasks: from trip planning and online shopping to composing emails and organizing work.

At the same time, the company is introducing a new optional subscription Comet Plus for $5 a month. In return, it offers access to premium content from leading publishers (like CNN, The Washington Post, or Wired), proposing a new revenue-sharing model rather than pay-per-click. Plans also include a mobile app and a “Background Assistants” feature — agents performing tasks in the background.

OpenAI did not intend to fall behind and presented ChatGPT Atlas. This is their vision of a browser where ChatGPT is no longer just a tool where we paste text, but a super assistant that understands the context of our work on the web.

Atlas introduces two key concepts: “browser memories” and “agent mode.”

Memories allow ChatGPT to remember the context from visited pages. This way, we can ask questions like: “Find all the job listings I viewed last week and prepare a summary of industry trends so I can get ready for interviews.” OpenAI assures that this feature is optional and users retain full control over what the assistant remembers.

This “agent mode” reveals the true power of LLMs. Atlas can not only answer questions but also perform tasks on our behalf. We can ask it to find a recipe, add all ingredients to the shopping cart, and order them for home delivery. In a business context, the agent can analyze old company documents, conduct competitor research, and prepare a brief for the team.

Atlas is currently available on macOS, with Windows and mobile versions expected soon.

The vision of a proactive assistant managing our digital lives is exciting but carries fundamental risks. The more permissions and data access AI has, the more dangerous security vulnerabilities become.

Shortly after Comet’s launch, security researchers at Brave (creators of a competing browser) discovered a critical security vulnerability (CVSS 8.6) in it. The issue concerns highly deceptive “prompt injection” attacks.

An attacker can embed on a website text (a malicious instruction for AI) almost the same color as the background. This text is invisible to humans, but when the user takes a screenshot, Comet’s OCR (optical character recognition) function reads everything — including the hidden command.

The problem was that the browser did not filter or validate text coming from OCR and treated it as a user command. Thus, an attacker could force the AI to perform unauthorized actions such as stealing banking data, exfiltrating emails, or taking over corporate systems.

Brave’s researchers emphasize that this is not an isolated issue of Perplexity. Similar vulnerabilities were found in other “agent browsers.” OpenAI, regarding Atlas, assures users have full control over their data, and content is not used by default to train models.

However, the fundamental problem remains. As George Chalhoub from the UCL Interaction Centre put it: “There will always be some risks related to prompt injection because that’s the nature of natural language interpreting systems. […] It’s a bit like a cat and mouse game.”

Until AI browsers learn to precisely distinguish trusted user commands from untrusted content originating from a webpage, we will remain exposed to new attack vectors.

On one hand, browsers integrated with LLMs can offer us incredible possibilities — AI agents who save our time and become real assistants. On the other hand, we see new serious security threats. There is also the question of habits. We are attached to our browsers, which we have been using for years. Will the convenience offered by Atlas and Comet be enough reason to abandon old habits? Time will tell, but one thing is certain — the race to create the definitive “AI agent” for web browsing has just begun. Who will win? If you’re curious about this issue, stay updated with AI news by subscribing to Delante’s newsletter!

Sources of information about AI browsers: