- The report warns the pirates exploit browser agents who do not know how to identify the false url
- A browser AI agent gave Google Drive access to a malicious application without hesitation
- Squarex says that AI agents are more vulnerable than humans with cyber attacks even basic
A spectacular change in business safety has appeared with the adoption of AI navigator agents, an automated tool that interacts with the web on behalf of users – but these agents have now become a major blind spot in cybersecurity defenses.
New research from Squarex has said that navigator AI agents are more likely to become the prey of cyber attacks than employees – contesting the long -standing belief that human error is the lowest link.
Unlike staff who are following regular cybersecurity training, agents cannot recognize “suspicious URL, excessive authorization requests or unusual websites”, explains the company.
A new, weakest link emerges in corporate cybersecurity
“The arrival of AI navigator agents dethroned employees as the weakest link in organizations,” said Vivek Ramachandran, CEO of Squarex.
These agents are able to imitate user behavior to perform tasks such as flight booking, meeting meetings or response to emails – however, their fundamental weakness lies in their total lack of security intuition.
Their responses are fully focused on tasks and devoid of the critical thinking necessary to assess risks.
In a significant demonstration, Squarex used the open source browser use framework to ask an AI agent to register for a file sharing tool.
Rather, the agent granted access to malicious application to the user’s messaging account, despite “unrelevant authorizations, unknown marks, suspect URLs” which would have stopped a human.
In another case, an agent was led to enter the connection identification information on a phishing site, following a dirty routine connection instruction.
Part of the danger stems from the way in which the agents of the AI of the browser work, because they work with the same privileges as the user, which makes their actions indistinguishable in legitimate behavior.
“Optimically, these agents have the awareness of the safety of an average employee, making them vulnerable to the most basic attacks, even less to those of bleeding,” said Squarex.
“In a critical way, these AI browser agents work on behalf of the user, with the same level of privilege to access business resources.”
Once an agent is compromised, the attackers have an access not detected to the internal systems, with all the authorizations of a trusted employee.
The current harvesting of security solutions, ranging from the best final point protection to the best ZTNA solution, does not take sufficient account of these agents.
Even the best Fwaas deployments can find it difficult to report actions that seem legitimate but come from a compromised AI.
“Until browsers develop native railings for AI agents of the browser, companies must incorporate native browser solutions such as the detection and response of the browser to prevent these agents from being deceived in malicious tasks,” note the researchers.
However, the wider message remains urgent: AI agents need not only intelligent but more intelligent supervision.
