New branch steering attacks threaten computer-use agents, COBRA architecture offers defense

Researchers have identified a new class of security vulnerabilities in computer-use agents (CUAs) called branch steering attacks. These attacks exploit the dynamic nature of CUAs, which must branch their execution paths based on anticipated runtime web content. Adversaries can craft untrusted data to guide the CUA down a hazardous, pre-approved branch without injecting explicit malicious instructions. A new benchmark, STEER-Bench, demonstrated high attack success rates against standard and Dual-LLM CUAs. To address this, a new architecture called COBRA was proposed, which pairs trusted branching plans with ahead-of-time capability constraints, significantly reducing attack success while maintaining high utility. AI

IMPACT Introduces a new attack vector for AI agents and a potential defense, impacting the security landscape of AI-driven automation.

RANK_REASON Academic paper detailing a new security vulnerability and proposed defense mechanism. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Cobra
• Computer Use Agents
• Planner LLM
• P-LLM
• Q-LLM
• Quarantined LLM
• STEER-Bench

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →