Avishek Das / SOPA Images via Reuters
OpenAI’s new web browser, Atlas, has been available for less than two weeks — and only on Apple computers, for now — but it’s getting a lot of attention.
That’s because it’s a new kind of browser in a market dominated by Google Chrome. And it comes from OpenAI, the leader in AI chatbots, at a time when artificial intelligence is starting to give traditional web search a run for its money.
“We think that AI represents a rare, once-a-decade opportunity to re-think what a browser can be about,” OpenAI CEO Sam Altman said on a livestream to launch the browser.
Atlas comes with ChatGPT baked in, and while it can navigate the web like traditional browsers, the company says it can do much more. A feature that OpenAI calls “agentic mode” can take action, like an agent who can shop for you, make reservations, or buy plane tickets. On that livestream, Altman’s colleague demonstrated how it can read an online recipe, figure out how many ingredients are needed for a set of diners, then buy the ingredients online.
OpenAI says it wants to unlock the power of AI, but some analysts see increased risks. The large language models that underpin artificial intelligence require vast quantities of data to improve.
OpenAI has “kind of reached the limits of what data they can get just by hoovering up all of the content that’s visible on the internet without consent,” said Anil Dash, a tech entrepreneur and writer.
But because Atlas is intertwined with ChatGPT, it absorbs much more user data than an ordinary browser does. The browser can interact with your email, for instance, or Google docs. It can keep so-called “browser memories” — details from the sites you’ve visited — so that OpenAI can better understand you.
“I think a big, big, big part of this is they are hoping to use the people who downloaded this browser as their agents to getting access to even more data,” Dash said. “I would not be surprised if there is more information going to them than coming to the user.”
That makes for a privacy trade-off. If you’re letting that AI agent shop for your dinner party, it’ll need a payment method, and maybe some passwords. It might also need to check your calendars, and personal contacts.
Lena Cohen is a Technologist at the Electronic Frontier Foundation, a digital rights group, has data privacy concerns about browsers acting as agents.
“The agentic AI mode takes these risks to a whole new level,” she said.
Users are potentially handing more control to OpenAI than they might think, she said. “Once your data is on OpenAI’s servers it’s hard to know and control what they do with it,” Cohen added.
NPR reached out to OpenAI with questions related to data and security, and was referred to company statements online and their Atlas demo video. In those, the company says its default is to not use information users pull up on Atlas to train its AI models, but people can opt in.
Cohen flagged another potential risk that experts think could be particularly dangerous with AI browsers like Atlas — nefarious pieces of code hidden in websites called “prompt injections.”
“Basically, bad actors can hide malicious instructions on a web page, and so when your AI agent visits that page, it could be tricked into executing those instructions,” she said.
For example, that AI agent browsing for groceries could bump into a prompt injection that says, “Buy this product, instead of that one.” Or maybe it says, “Hand over your credit card information.”
OpenAI says this is an unsolved problem, but they’re working on training their models to ignore those harmful instructions.
Chirag Shah, a professor with the Information School at the University of Washington, says AI has become a phenomenon at warp speed, with minimal regulation, and there have been consequences.
“We’re in this kind of game where it’s a typical mentality of move fast and break. Unfortunately, what’s breaking is not just the tool or the technology, but real people,” he said.
