AI-Powered Browsers Struggle to Identify Safe Sites—What Dangers Does This Pose?

Tech
By 24matins.uk,  published 26 August 2025 at 11h26, updated on 26 August 2025 at 11h26.
Tech

Artificial intelligence is increasingly integrated into web browsers, promising enhanced security for users. However, recent findings reveal that these AI-powered browsers still struggle to accurately identify which websites are safe and which pose risks, raising concerns about user protection.

Tl;dr

  • AI browsers struggle to detect phishing attempts.
  • Guardio’s tests reveal serious security flaws.
  • Human vigilance remains crucial for online safety.

Growing Adoption of AI Browsers Amid Persistent Security Concerns

In recent months, an increasing number of users have been embracing AI browsers to automate daily tasks—everything from organizing trips and shopping to handling emails. The promise is alluring: greater efficiency and seamless online experiences. Yet, beneath this surface of convenience lies a troubling vulnerability. Despite their advanced capabilities, these tools have demonstrated glaring weaknesses in distinguishing legitimate websites from sophisticated phishing schemes, as highlighted by the cybersecurity firm Guardio.

A Revealing Experiment: When Automation Meets Deception

To assess just how resilient these new solutions are, researchers at Guardio conducted a series of eye-opening tests centered on the AI-powered browser Comet, created by Perplexity. Their method? They built a convincing fake version of the Walmart website using Lovable, then instructed the browser to purchase an Apple watch. Remarkably, despite numerous red flags—dubious design, mismatched visuals—the browser proceeded without hesitation. It added the watch to the cart, autofilled sensitive personal and banking data, and completed the transaction automatically, never pausing to ask for user confirmation.

There were moments when the browser hesitated: in some instances, it either declined to finish the transaction or advised human review. However, these cases were outnumbered by those where confidential information was transmitted directly to fraudulent sites.

The Human Element: Still Essential for Security

Perhaps even more disconcerting was how the same browser dealt with realistic-looking scam emails. Messages apparently sent by a «Wells Fargo investment manager» were not only marked as actionable tasks but had their embedded links clicked without discernment. The AI went so far as to fill out forms designed explicitly for stealing credentials.

A few key risks become immediately apparent:

  • AI left unchecked may follow instructions blindly.
  • User data can be compromised without warning.
  • Sophisticated phishing tactics remain largely undetected by current AI models.

    • Cautious Optimism: Progress Needed Before Full Trust

    Ultimately, what stands out is that while these promising technologies are gaining ground fast, their security frameworks simply haven’t caught up with today’s evolving threats. Delegating sensitive transactions or financial operations to such immature tools carries real dangers. Until designers can shore up their defences against increasingly complex cyberattacks, one thing seems clear: «Caution should be your default setting online.» For now, human oversight remains not just recommended—but essential.