OpenAI’s new browser, ChatGPT Atlas, blurs the boundary between tool and companion. Announced this week, it represents a significant step toward a web that does not just display information but interprets, remembers, and acts on it. Yet behind the convenience lies a growing unease among cybersecurity experts who see Atlas as a window not just to the internet, but into the user’s mind.
Unlike traditional browsers, Atlas integrates ChatGPT directly into the browsing experience, allowing the AI to understand webpages, recall past activity, and perform actions such as booking appointments or compiling research. The result is a seamless, hyper-personalised experience, an AI that travels with users across the web, learning from their habits, preferences, and goals.
But this same intimacy has triggered alarm. Atlas introduces two features, browser memories and agent mode, that give the AI permission to store contextual details from browsing sessions and take actions on the user’s behalf. Critics say this transforms the browser into a potential surveillance tool, capable of creating a detailed behavioural map that extends far beyond search history.
When convenience becomes surveillance
“Search has always been surveillance. AI search made it intimate surveillance,” said Eamonn Maguire, Director of Engineering for AI and ML at Proton. “OpenAI’s new ChatGPT Atlas takes another step, total surveillance. Traditional search engines see isolated queries, but conversational AI captures context and narrative. It learns not just what you look for, but how you think.”
Maguire warned that Atlas could “track everywhere you go, what you think, want, and feel,” arguing that users will inevitably share details about their health, finances, and personal lives in ways they never would with a search bar. “The result is surveillance capitalism’s final form, AI so human-feeling that users willingly divulge intimate details while handing over browser-level access to their digital life.”
OpenAI has stressed that data sharing remains under user control and that browser memories are optional. Users can delete history or disable visibility on any page, and browsing data is not used for model training unless users explicitly opt in. But critics point out that such controls rely on trust, and on users understanding the trade-offs between convenience and exposure.
Dray Agha, Senior Manager of Security Operations at Huntress, described the launch as a “classic security dilemma: trading privacy for convenience.” He warned that while OpenAI has promised not to train models on user data by default, similar promises in the past have proven inconsistent. “Atlas introduces a rich, private profile of your online activity,” Agha said. “The privacy implications of how these ‘memories’ are stored, processed, and protected remain a key area for scrutiny.”
The rise of agentic browsing
For cybersecurity experts, Atlas marks the beginning of what they call agentic browsing, a shift from passive interaction to delegation, where users ask AI not just to find information but to act on it. In Atlas’s ‘agent mode’, ChatGPT can open tabs, summarise pages, or even make purchases. While the system pauses for user confirmation on sensitive sites, experts say the potential for misuse remains.
“Features like memory and agentic behaviour mean these systems can learn about you over time, build detailed profiles, and even act on your behalf,” said Charlotte Wilson, Head of Enterprise at Check Point Software. “That’s convenient, until it isn’t.” Wilson noted that such capabilities make AI-assisted browsers vulnerable to new forms of attack, where malicious websites could issue hidden instructions to manipulate the agent or extract data. “The technical risks are just as real. Modern exploits no longer need code, they can use language,” she said.
The issue extends beyond security to the integrity of information. Wilson cautioned that some large language models have been exposed to propaganda or biased datasets, raising questions about whether AI-assisted browsing could inadvertently amplify disinformation. “AI-powered search may feel more human and helpful than Google,” she said, “but it comes with a different kind of danger.”
A turning point for digital trust
For many analysts, Atlas is less about the technology itself than the cultural moment it represents. It collapses search, browsing, and reasoning into a single AI-driven interface, eliminating the boundaries between human intention and machine execution. That integration could redefine the way people navigate information, but also how their data is used.
“Consolidating browsing, reasoning, and summarisation into one AI window sounds great and convenient,” said Javvad Malik, Lead CISO Advisor at KnowBe4, “but it also aggregates risk into one point. People will inevitably run sensitive queries, upload documents, and click through assuming it’s just regular browsing, when it isn’t.” Malik urged users to treat AI not as an index but as a data processor, and to separate personal and professional use.
Experts agree that the future of AI-assisted browsing depends on transparency and education. “The human-AI interaction layer is where the real battle for security will be fought,” said Dr Martin Kraemer, CISO Advisor at KnowBe4. “We need better governance, literacy, and clear policies to ensure responsible adoption.”
ChatGPT Atlas may indeed redefine how people use the web. But it also forces a more difficult reckoning, how much autonomy and privacy users are willing to surrender in exchange for intelligence that feels indistinguishable from their own.
