MARK Suite logo
§ Governance

Just nine months after launching its browser Atlas, OpenAI discontinued working on it.

By Malin Gaertig · August 16, 2026

OpenAI is retiring Atlas as a standalone browser and folding its agent features into ChatGPT plus a Chrome extension. For brands, the browser layer is where agents watch, click and buy across any site — and it's where a new fraud and liability gap is opening.

Share
§ Takeaways
  • —OpenAI is discontinuing Atlas as a standalone browser; the browser-agent layer is consolidating around extensions and built-in browser features rather than standalone apps.
  • —Perplexity's Comet has momentum, but Gemini's auto browse in Chrome may be the most consequential because it requires no download and ships with the browser most people already use.
  • —Agentic browsers can bypass same-origin protections, and when they transact they look like the legitimate customer — creating a fraud and liability gap security teams should review now.

Just nine months after launching its browser Atlas, OpenAI discontinued working on it. Instead they'll fold the agent features into the ChatGPT app and have a Chrome extension instead.

This is the fourth in a short series on the different types of agents shaping how, and whether, brands get chosen. The first three looked at the model apps themselves, the closed marketplaces, and third parties building inside someone else's AI. This one is different again: agents that operate at the level of the browser, watching, clicking and buying across whatever site a person visits, logged in as that person.

Perplexity's Comet has the clearest momentum, crossing 10 million users within months of going free in October 2025 and available across every major platform. It's also the one Amazon sued over unauthorised shopping access and lost on appeal.

Atlas was OpenAI's attempt to own the browser layer outright rather than depend on Chrome, Safari and Edge for distribution. Retiring it as a standalone product nine months in suggests that it wasn't commercially viable.

The one worth watching is Gemini's auto browse built directly into Chrome itself rather than requiring a download. It will simply already be in the browser most of the world already uses.

Claude's version, by contrast, is a Chrome extension aimed more at work tasks than consumer shopping, consistent with the enterprise lean we've seen from Claude throughout this series. Microsoft has folded Edge's own Copilot features into the free browser as well.

University of Washington researchers tested seven agentic browsers and found four, including Atlas, Chrome with Gemini, Claude's extension and Comet, could be tricked into bypassing the same-origin policy, the rule that's kept one website from reading another's data for three decades. They ran a working attack against Atlas.

From a business perspective that leaves them exposed because when an agent buys on a customer's behalf, it looks exactly like that customer, logged in, on their own device. Your fraud tools are built to catch the wrong person in the right session, not the right session hijacked by instructions the agent picked up elsewhere. That's a live fraud and liability gap, worth a conversation with your security team now, not after the first incident.

§ Briefing

Just in: What Australia's Trust Gap in Agentic Commerce Means for Brands Selling Through AI Agents

Every week now brings a fresh batch of proof that agentic commerce is moving from slide-deck concept to live infrastructure. This week it was five stories at once: a landmark ad tech ruling, the biggest layoff in a category-leading DSP's history, an acquihire that reveals what platforms actually value in AI, a data point out of Australia that every retailer should sit with, and a payments infrastructure race that's outpacing the shopping experiences it's meant to serve. None of these stories are really separate. Read together, they say something specific about where the next twelve months are headed, and it's not the story most agentic commerce hype has been telling.

Read →
§ Briefing

Just in: What the EU's ChatGPT Ruling Means for Brands Selling Through AI Agents

Regulators, mall operators, measurement vendors and enterprise software giants are all independently racing to set the rules for how AI agents discover, compare and transact on our behalf. None of them are waiting for the rest of us to catch up, and none of them are coordinating with each other. If you are selling anything through an AI agent, in any market, that gap between fast-moving infrastructure and slow-moving standards is where your risk and your opportunity both live right now.

Read →
§ Briefing

Just in: What ChatGPT Selling Ads Means for Brands Selling Through AI Agents

Something quiet happened yesterday that will shape how your brand gets found for the next decade. On 24 August, OpenAI switched on advertising inside ChatGPT across 31 European markets. On its own that reads like a monetisation story. Sat next to a courtroom decision on shopping bots, a wave of agent-standards announcements, and a measurement industry cutting jobs by the hundred, it reads like something bigger: the buying journey is migrating to a layer most brands still cannot see, let alone influence. Here is why this week matters, and what senior leaders can actually do about it.

Read →
§ Briefing

Just in: What Google Killing Assistant for Gemini Means for Brands Selling Through AI Agents

Three separate stories broke this week, in three unrelated corners of the industry, and they are all describing the same shift from different angles. Google switched on Gemini as the default AI across billions of Android devices. A major payments study found that most merchants cannot see AI agents shopping on their sites even as they happen. And a US appeals court quietly decided who is legally responsible when an agent, not a person, clicks buy. Read together, they are the clearest signal yet that agentic commerce has stopped being a 2027 planning slide and started being a this-quarter operating problem.

Read →

Go deeper: Product suite · Framework guides.