Users of browser AI agents may be vulnerable to significant cybersecurity risks, says new research from the University of Washington.
Through successful proof-of-concept cyberattacks, researchers bypassed a fundamental security measure, the same-origin policy, a protocol that usually keeps information siloed between pages. While humans may be savvy enough to not be tricked, certain types of attacks may make AI agents vulnerable to sharing private credentials and information.
As of now, researchers have not found a solution to mitigate the user risks exposed while maintaining full browser capabilities. After decades of building up the same-origin policy, the rise of browser AI agents has meant big steps back for browser security.