Browser agents are designed to make decisions and execute actions on behalf of users, but they rely on untrusted webpages as input. That opens the door to prompt injection: malicious instructions embedded in a page that can hijack a legitimate task. A new benchmark, BrowseSafe++, is being introduced to study how well these agents detect such attacks.
According to the arXiv abstract, BrowseSafe++ spans multiple dimensions that matter in real browsing: different web contexts, different languages, and different browser views. This suggests the benchmark aims to capture the variety of situations where a browser agent might encounter hostile content, rather than testing only a narrow set of cases.
The abstract is brief, so details on the benchmark's construction and results are not yet available. Still, the work highlights a growing concern: as browser agents become more autonomous, defending them against injected instructions is as important as their ability to complete tasks.