OpenAI switched off its ChatGPT Atlas browser on Sunday, 9 August, less than 10 months after it launched. The tools that made Atlas an “agentic” browser — opening tabs, logging into sites, downloading files and letting an AI agent run tasks for you — now live inside ChatGPT and OpenAI’s Codex coding tool instead.
The company announced the shutdown in July, in its release notes, next to a new desktop app called ChatGPT Work that folds ChatGPT, Codex and Atlas into one program. Atlas first appeared on 21 October 2025, for Mac computers, and was built on Chromium, the same open-source engine behind Google Chrome. Users got roughly a month’s notice to move off it.
Atlas’s abilities are being split across two browsing surfaces in the new app. One is a built-in browser that can visit sites, sign into accounts and save files. The other is a cloud browser that runs on OpenAI’s own servers, where an agent works through a task remotely while you do something else. OpenAI also released a ChatGPT extension for Chrome.
The switch-off came with homework. Atlas bookmarks do not move over on their own, so OpenAI told users to export them as an HTML file and save any pages they wanted to keep before the deadline. Open tabs and browsing history stayed behind too. ChatGPT chat history is stored separately and survives. OpenAI said it did not want people stuck on a discontinued browser that could stop getting security updates.
The reason for the retirement is a wider clean-up at the company. Fidji Simo, OpenAI’s former head of applications, had told staff to cut “side quests,” and the same push shut down the Sora video tool earlier in 2026. James Sun, who leads OpenAI’s browsing work, framed Atlas as a test bed rather than a failure. “All these capabilities were built on what we learned from Atlas users,” he wrote on X, adding that OpenAI was “applying these learnings to these new products.”
Beyond the timeline, the short life of Atlas says something about how fast this category is moving. The AI browser race is barely a year old and already crowded, with Perplexity’s Comet and Anthropic’s Claude extension chasing the same idea. It also carries a safety problem no one has solved: researchers recently tricked six AI browsers, Atlas among them, into leaking user credentials through hidden instructions on web pages, an attack known as prompt injection.
Independent developer Simon Willison, who coined the term, has warned about exactly this setup. He calls it the “lethal trifecta”: an agent that can see private data, read untrusted web content and send data out can be turned against its owner. Even after a rare good experience with a browser agent in December 2025, he wrote that he remained “deeply skeptical of the entire browsing agent category.”
OpenAI is not walking away from browsers. It is betting the browser belongs inside ChatGPT, and that the assistant becomes the place people work. For users, the lesson is blunter: OpenAI’s newer experiments can disappear quickly, so anything built on top of them is worth backing up. The open question is whether pushing agents deeper into ChatGPT, where they can act inside your logged-in accounts, makes that unsolved security risk bigger rather than smaller.
