Browser AI agents identified as vulnerable to phishing and OAuth attacks
SquareX research demonstrates that Browser AI Agents are vulnerable to phishing and OAuth attacks, reportedly more susceptible to compromise than human employees. A demonstration using the popular Browser Use framework showed an agent granting unauthorized access to a user's email and other data despite suspicious warning signs.
What happened
SquareX research demonstrates that Browser AI Agents using the open-source Browser Use framework are vulnerable to OAuth attacks and phishing. A Browser AI Agent instructed to find and register for a file-sharing tool granted a malicious app complete access to a user's email and Google Drive despite multiple suspicious signals including irrelevant permissions, unfamiliar brands, and suspicious URLs.
Reported impact
Browser AI Agents are reportedly more vulnerable than human employees to cyberattacks and lack ability to recognize visual warning signs like suspicious URLs, excessive permission requests, or unusual website designs. The agents could expose users' credit card information to phishing sites or disclose sensitive data when responding to impersonation emails.
Index analysis
Unlike human employees, Browser AI Agents are not subject to regular security awareness training and cannot recognize basic visual warning signs of malicious websites. Creating security guardrails for every task would require extensive prompts that would likely outweigh productivity gains, and employees lack the security expertise to write such prompts.
Governance controls
Traditional browsers and security tools cannot differentiate between actions performed by users and those performed by agents. Enterprises require browser-native guardrails to prevent agents from performing unauthorized actions.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.