AI Browsers: The Unseen Security Risk and Industry Shifts You Need to Know

As the realm of artificial intelligence expands rapidly, we're beginning to see new applications of AI technologies in our everyday tasks. Enter AI browsers, like Fellou and Comet from Perplexity, which are presenting themselves as the next evolution in web browsing. With features that enable them to read and summarize web pages, these intelligent tools promise us efficiency and speed in our online research or digital workflows. But, here’s where things get tricky—security experts are waving red flags, signaling that the rise of these AI browsers could pose unforeseen risks to enterprises.

Are AI Browsers an Invitation for Trouble?

The underlying issue is that AI browsers are susceptible to indirect prompt injection attacks. What does that even mean? Essentially, it’s when the AI browser encounters sneaky instructions hidden in cleverly crafted websites. These instructions can be embedded in texts or images, cleverly disguised to slip past human detection. Once the AI model picks these up, it may unknowingly perform actions that could jeopardize sensitive information or company data.

So, what’s the takeaway for IT departments? Right now, the consensus is clear: AI browsers are not yet safe for enterprise use, representing a significant security threat.

Autonomy Meets Vulnerability

Tests have shown that these AI browsers can decode harmful embedded texts as user instructions, which can then be executed with the user's privileges. Imagine this—if a user has access to sensitive corporate data, the risks multiply. With increased autonomy comes a broadened attack surface, heightening the chances of data loss. An example would be if a malicious actor embeds text commands within an image, which, upon display, could trigger the browser to access corporate emails or banking information without user approval.

This introduces shadow AI—unauthorized AI activity that acts outside of company protocols and guidelines. Essentially, these AI models bridge data from various domains while bypassing security measures meant to keep information compartmentalized.

Challenges in Governance and Implementation

The core problem stems from how user queries in a browser intertwine with live data accessed online. If the Large Language Model (LLM) used in the AI browser fails to differentiate between safe user inputs and harmful instructions, it could unwittingly act on unauthorized data. This raises a serious alarm for organizations relying on strict data segmentation and access control.

The real kicker? If a user’s AI browser is compromised, it can mimic the user’s behavior to access protected information, making the AI browser akin to an insider threat. Users may remain blissfully unaware as their compromised browsers operate unnoticed for extended periods.

Strategies for Threat Mitigation

For any robust IT security team, the first generation of AI browsers ought to be treated like software that’s been installed without authorization. It’s one thing to restrict what software users can install, but it’s becoming increasingly clear that even mainstream browsers like Chrome and Edge are integrating more AI features, such as Gemini and Copilot, intensifying this challenge. Large browser companies are competing fiercely to enhance their offerings, often at the expense of security!

As we roll into the next generation of browsers, organizations must check for specific capabilities:

• Prompt isolation, which clearly distinguishes user intent from third-party web content
• Gated permissions, ensuring AI agents can’t act autonomously without user confirmation
• Sandboxing sensitive browsing activities to prevent AI operation in critical areas
• Governance integration with strong compliance to data security policies, enabling traceability for all agentic actions

To date, no browser manufacturer has developed a system capable of clearly separating user-driven requests from AI-interpreted commands. This means the potential for simple prompt injection to lead to harmful actions is still all too real.

The Bottom Line for Decision-Makers

Agentic AI browsers are being marketed as the next big thing in web operations and automation, expertly blurring the line between human activity and smart tech interaction. However, given the alarming ease with which current AI browsers can be manipulated, it's prudent to consider them dormant malware.

As major browser vendors contemplate AI integration—agentic features included—it’s essential that careful oversight is exercised with each new rollout. After all, when it comes to safeguarding enterprise data, we can't afford to overlook the details.