
Significant Agentic AI browser vulnerabilities have surfaced, which could compromise millions of users simultaneously. An analysis of autonomous AI browsing systems reveals that they introduce unprecedented AI security risks that traditional protection mechanisms cannot adequately address.
Guard.io's comprehensive Scamlexity report reveals that the research team tested Perplexity's Comet browser using three distinct attack scenarios, demonstrating how AI agents fail to apply human-like skepticism when encountering malicious content.
In controlled experiments, the AI browser autonomously completed purchases on fraudulent e-commerce sites, processed phishing emails without verification, and fell victim to novel "PromptFix" attacks that exploit AI-specific vulnerabilities.
The most concerning finding involved automated transaction completion on fake shopping sites.
When prompted to "buy an Apple Watch," the AI agent navigated through a convincing but fraudulent Walmart replica created via Lovable AI, autofilled saved payment credentials, and completed the purchase without human oversight or verification protocols.
The research identified a sophisticated attack methodology called PromptFix, representing an evolution of traditional ClickFix scams adapted for AI systems.
This technique embeds hidden prompt injection instructions within legitimate-looking CAPTCHA forms, causing AI browsers to execute malicious commands while appearing to assist users in solving authentication challenges.
These Agentic AI browser vulnerabilities demonstrate how attackers can bypass human intuition entirely, targeting the AI's core programming to assist users without questioning task legitimacy.
The trust chain corruption occurs when AI agents process malicious instructions embedded in webpage content, leading to unauthorized downloads, credential exposure, and system compromise.
“Before the arrival of GenAI, attackers were already proficient at rapidly creating new domains to bypass traditional phishing detection tools,” says Krishna Vishnubhotla, Vice President, Product Strategy at Zimperium.
“Sophistication shows up in the form of highly realistic and personalized, well-written phishing content at scale across all mobile phishing (mishing) vectors, including audio, video, and voicemail. The automation aspect allows attackers to clone websites in seconds, making brand impersonation easier than ever.”
Guard.io's analysis reveals that current security frameworks remain insufficient for protecting autonomous AI browsing systems. Traditional safeguards like Google Safe Browsing failed to prevent multiple attack scenarios, highlighting the need for AI-specific security architectures.
The cybersecurity firm recommends implementing comprehensive guardrails within AI decision loops, including enhanced URL reputation checking, behavioral anomaly detection, and mandatory human verification for sensitive operations.
Organizations deploying Agentic AI browsers must integrate security considerations into system architecture rather than treating protection as an auxiliary component.
“Specific types of AI can perform thousands of calculations in real time to detect suspicious behavior and perform the micro decision-making necessary to respond to and contain malicious behavior in seconds,“ says Nicole Carignan, Senior Vice President, Security & AI Strategy, and Field CISO at Darktrace.
“Transparency and explainability in the AI outcomes are critical to foster a productive human-AI partnership.”