Repository navigation
[new-research] arXiv:2610.03089 — Securing Computer-Use Agents Against Branch Steering Attacks #198
Description
New arXiv Research Paper
arXiv ID: 2610.03089
Title: Securing Computer-Use Agents Against Branch Steering Attacks
Authors: Giulio Zingrillo, Hanna Foerster, Ilia Shumailov, Yiren Zhao, Robert Mullins
Published: 2026-10-02T10:08:39Z
URL: https://arxiv.org/abs/2610.03089
Abstract (first 300 chars)
Modern Computer Use Agents (CUAs) directly interact with graphical user interfaces and execute third-party web tools, exposing them to indirect prompt injection across every rendered page and tool response. While the Dual-LLM pattern is the primary system-level architecture offering formal security …
Suggested OWASP Mapping
LLM01
Suggested Action
Review this paper against the OWASP GenAI Crosswalk entries and update relevant mapping files if new attack patterns or mitigations are identified.
Activity
emmanuelgjr commentedon Oct 9, 2026
Routed to the benchmark catalogue — rule 3.
Securing Computer-Use Agents Against Branch Steering Attacks studies branch steering against computer-use agents, where crafted data coerces an agent down a hazardous, pre-approved branch. It introduces STEER-Bench, "101 tasks across 9 domains", and proposes the COBRA architecture as a defense.
Routed per
docs/TRIAGE_RULES.md:evals/EXTERNAL_BENCHMARKS.md. COBRA is literature and is noted.The row's OWASP entry column lands as DRAFT — SME review required. This issue closes when #222 merges. A row for a benchmark introduced by a defense paper follows the maintainer ruling of 2026-10-09, recorded in rule 3 by that PR.