Agent interfaces leave 91% of tested web flaws exploitable
Original title: The Agent in the Middle: How AI-Mediated Access Reshapes Web Vulnerability Exploitation
VELA covers 44 vulnerabilities in twelve web families; interface design shifted exploit retention from 39% to 95%.
Researchers built VELA, an executable benchmark of 44 vulnerabilities across twelve web application families, with agent-facing interfaces derived from legitimate workflows. Three commercial agents realized every attack that direct interface calls realized in the reference setup, where 91% of flaws stayed exploitable. Removing operations or limiting input stopped some attacks but also constrained legitimate workflows. Across configurations, interface design moved retention from 39% to 95%, against at most 4.5 points of difference across agents.
Unconfirmed: The commercial agents are not named, and results apply only to the tested configurations.
- No reference-configuration miss was blamed on refusal or input normalization.
- In one case the agent supplied calls the attacker had omitted.
- Independent oracles verified whether attacks succeeded.
- Some restrictive interfaces also limit legitimate workflows.
From the original article
1 min read at arxiv.orgWeb services increasingly expose backend interfaces that let users delegate tasks to AI agents, which invoke service operations directly without navigating websites. This shift changes how existing Web vulnerabilities can be exploited. We investigate these changes through VELA, an executable benchmark covering 44 vulnerabilities across twelve Web application families.
By Ruozhao Yang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.