OpenAI's Atlas Browser Integrates ChatGPT, Raises Data Privacy Questions

OpenAI Atlas Browser Integrates ChatGPT, Raises Data Privacy Questions

OpenAI has released its new web browser, Atlas, currently available for Apple computers. The browser incorporates OpenAI's ChatGPT technology.

OpenAI CEO Sam Altman stated during the browser's launch that artificial intelligence provides an opportunity to redefine browser functionality. Atlas is designed to perform advanced tasks in addition to traditional web navigation, leveraging its integrated ChatGPT. A feature called "agentic mode" enables the browser to execute actions such as shopping, making reservations, or purchasing tickets. A demonstration during the launch event illustrated the browser's capability to process an online recipe, determine required ingredient quantities for a specified number of diners, and facilitate the online procurement of those ingredients.

The integration of ChatGPT within Atlas involves the browser collecting user data, which may exceed the data collection scope of standard web browsers. Atlas can interact with user services such as email and Google Docs, and it stores "browser memories," which include details from visited websites. OpenAI has indicated that this data collection is intended to enhance user understanding.

Anil Dash, a technology entrepreneur and writer, commented on these data collection practices, suggesting that OpenAI aims to access additional data through Atlas users. He expressed a belief that more information might be transmitted to OpenAI than is evident to the user.

Lena Cohen, a Technologist at the Electronic Frontier Foundation (EFF), has articulated data privacy concerns regarding browsers that function as agents. Cohen stated that the agentic AI mode significantly escalates privacy risks, noting that users may transfer more control to OpenAI than anticipated. She also highlighted the challenges in understanding and controlling data once it resides on OpenAI's servers.

OpenAI, in response to inquiries regarding data and security, referenced company statements and demo videos. These resources indicate that the default setting for Atlas is not to use user information for training its AI models; however, users have the option to opt-in to this practice.

Cohen also identified "prompt injections" as a potential security risk within AI browsers. This involves malicious instructions embedded in web pages that an AI agent could be manipulated into executing. Examples provided include an agent being instructed to purchase an unintended product or to disclose credit card information.

OpenAI acknowledges prompt injection as an unresolved issue but has stated that efforts are underway to train models to disregard such harmful instructions.

Chirag Shah, a professor at the University of Washington's Information School, has commented on the rapid advancement of AI with limited regulatory oversight.