ANGLER, an adaptive attack framework evaluated on eight vision-language web agents, achieved a 20.36% end-to-end attack success rate, compared with 15.24% in a non-adaptive setting. On the same benign tasks, attacks generated by ANGLER reduced legitimate task completion from 61.9% to 44.4%.
The study introduces REEF, a security benchmark covering nine websites across travel, commerce, productivity, media and communication. It pairs legitimate tasks with malicious objectives using five attack types at the user and environment levels. State-based verifiers check actual changes, rather than judging attack text alone. ANGLER uses feedback from earlier runs to mutate attacks and select candidates, prioritizing near misses. The results show how adaptive adversarial content can both trigger unauthorized actions and disrupt useful work in these test environments.


arXiv.org
ANGLER: Verifier-Grounded Benchmarking with an Adaptive Attacker for Web Agents
Web agents are increasingly deployed to complete tasks autonomously on real websites, where adversarially injected content can induce them to perfo...

















