AI Browsers Pose Security Risks: Why I Can No Longer Ignore the Threat

In recent years, I’ve experimented with various web browsers, including the innovative AI-driven browsers known as agentic browsers. These platforms not only display websites but also operate autonomously in response to user requests. Currently, some of the most recognized AI browsers are Perplexity’s Comet and Dia, alongside Google’s incorporation of Gemini within Chrome. While these AI browsers can streamline certain online tasks, they also raise significant concerns regarding data security and privacy.

Understanding AI Browsers

AI browsers are essentially web browsers enhanced with artificial intelligence capabilities that go beyond simple website navigation. Unlike traditional browsers that merely display content, AI browsers act as proactive assistants, ready to address your queries and perform tasks on your behalf.

Commonly referred to as agentic browsers, they possess the ability to interpret user intentions, traverse websites independently, and complete various tasks, such as filling out forms, booking reservations, shopping online, summarizing articles, and managing email and calendar applications. Instead of the typical clicking and scrolling, users can simply issue commands like “Book dinner for Friday, ” and the AI will manage the entire booking process.

Among the notable players in this arena, Perplexity’s Comet stands out for being a fully integrated browser built around agentic AI. Designed to operate like a personal assistant, Comet excels at summarizing web pages, organizing inboxes, and executing complex workflows.

Built on the Chromium framework, Comet supports natural language commands and seamlessly integrates with services such as Gmail and Calendar, although it is currently available only through a premium “Max” plan priced at $200 per month. I tested Comet myself and found its potential promising, yet I ultimately reverted to my standard browser due to concerns regarding privacy.

Dia, created by The Browser Company and now part of Atlassian, provides a comparable experience, boasting overlapping capabilities with Comet.

While OpenAI hasn’t launched a distinct AI browser, its ChatGPT application offers an Agent Mode capable of tackling similar functions. Furthermore, Google is venturing into this domain with Project Mariner, a Chrome extension designed to facilitate shopping, information retrieval, and form completion.

Data Privacy Risks of AI Browsers

AI-driven browsers and agents require extensive access to user data, which can trigger significant privacy concerns. These tools often amass substantial amounts of browsing history and user interactions for AI processing. The functionalities that enhance the efficiency of these browsers—such as reading multiple tabs, indexing local files, and integrating with email or calendar systems—mean they could inadvertently process sensitive user data without adequate safeguards in place.

Shivan Kaul Sahib, VP of Privacy and Security at Brave, provided insight on security implications: “The security risk depends on how exactly AI is integrated into the browser… When AI is an assistant with capabilities that are tightly controlled, the security risk mirrors that of traditional browsing. However, giving AI the ability to perform actions autonomously complicates security dynamics.”

Further highlighting potential vulnerabilities, researchers at Brave found a critical risk associated with Comet. An experiment allowed cross-domain account access via malicious prompts that instructed the AI assistant to summarize a webpage. The AI inadvertently processed content by blending user commands with the web page text, which could potentially allow attackers to exploit this functionality by embedding deceptive instructions within a page.

Moreover, security experts warn that AI agents increase the “attack surface” for cyber intrusions. Sam Altman, CEO of OpenAI, acknowledged that granting control to an AI agent enhances risks and expands the attack vector substantially.

Weighing Convenience Against Privacy

While AI browsers promise to simplify online activities, they come with significant privacy trade-offs. Through my experience testing these platforms, I observed that their convenience often undermined speed, reliability, and security. For example, when using Comet, I initially appreciated its potential to automate browsing, but I quickly grew frustrated as it took much longer to complete tasks compared to traditional browsing methods.

Similarly, my tests with OpenAI’s Operator yielded slightly better results but were far from flawless. Frequent mistakes led me to question its outputs rather than trust its recommendations.

Mitigating the Misuse of AI Browsers

Beyond performance concerns, the primary risk associated with AI browsers is their potential for misuse. A concerning experiment by security firm Guardio, dubbed “Scamlexity, ” demonstrated just how vulnerable these systems can be. In this study, researchers created a fraudulent e-commerce site and instructed Comet to purchase an Apple Watch. Alarmingly, in various trials, Comet proceeded to add the item to the cart and attempted checkout using the user’s saved credit card information without seeking confirmation.

Another experiment involved sending a fake phishing email from Wells Fargo. Comet, having access to the user’s inbox, not only opened the email but also clicked on the phishing link and began attempting to input login credentials on the spoofed website.

“AI-powered browsers represent a double-edged sword, ” cautioned Chandrasekhar Bilugu, CTO of SureShield. He emphasized the need for developers to implement safety protocols such as sandboxing, restricting AI access to sensitive features, and conducting adversarial testing. Users also bear responsibility by activating privacy extensions, refraining from disclosing sensitive information to AI tools, and insisting on transparency in AI processes to mitigate exploitation risks.

These insights underscore a critical issue: the very attributes designed to enhance user experience in AI browsers may simultaneously compromise privacy and security. Without substantial protective measures, utilizing AI browsers often means relinquishing control over sensitive data for underdeveloped functionalities.

For those interested in differentiating between AI and human interactions, several tools are available to help.

Source & Images