Images Avishek Das/SOPA 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 type 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 believe AI represents a rare, once-in-a-decade opportunity to rethink what a browser can be,” Sam Altman, CEO of OpenAI, said during a livestream for the browser’s launch.
Atlas comes with ChatGPT built-in, and while it can browse the web like traditional browsers, the company says it can do much more. A feature OpenAI calls “agent mode” can act, like an agent who can shop for you, make reservations or buy plane tickets. During this live stream, Altman’s colleague demonstrated how to read a recipe online, determine the number of ingredients needed for a set of diners, and then purchase the ingredients online.
OpenAI says it wants to unleash the power of AI, but some analysts see increased risks. The large language models that underpin artificial intelligence require large amounts of data to improve.
OpenAI has “kind of reached the limits of the data they can get by just sucking up all the visible content on the Internet without consent,” said Anil Dash, a technology entrepreneur and writer.
But because Atlas is closely related to ChatGPT, it absorbs much more user data than a regular browser. The browser can interact with your email, for example, or with Google documents. It may remember what are called “browser memories” (details of the sites you have visited) so that OpenAI can understand you better.
“I think a big part of this is that they’re hoping to use people who downloaded this browser as agents to access even more data,” Dash said. “I wouldn’t be surprised if more information reaches them than reaches the user.”
This constitutes a compromise when it comes to privacy. If you let this AI agent shop for your dinner, it’ll need a payment method and maybe a few passwords. You may also need to check your calendars and personal contacts.
Lena Cohen is a technologist at the Electronic Frontier Foundation, a digital rights group, and has concerns about data privacy regarding browsers acting as agents.
“Agentic AI mode takes these risks to a whole new level,” she said.
Users are potentially handing over more control to OpenAI than they realize, she said. “Once your data is on OpenAI’s servers, it’s difficult to know and control what they do with it,” Cohen added.
NPR contacted OpenAI with questions related to data and security, and was referred to the company’s statements online and their Atlas demo video. In them, the company says its default is not to use the information users extract on Atlas to train its AI models, but that people can opt-in.
Cohen flagged another potential risk that experts say could be particularly dangerous with AI browsers like Atlas: harmful pieces of code hidden in websites called “quick 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 carrying out those instructions,” she said.
For example, this AI agent who is shopping could come across a quick injection saying: “Buy this product instead of that one”. Or maybe it says, “Hand over your credit card information.”
OpenAI says this is an unsolved issue, but they are working on training their models to ignore these harmful instructions.
Chirag Shah, a professor at the University of Washington’s School of Information, says AI has become a phenomenon at lightning speed, with minimal regulation, and there have been consequences.
“We’re in this kind of game where the typical mentality is to go fast and break. Unfortunately, what breaks is not just the tool or the technology, but real people,” he said.
