AI RESEARCH
The Autonomy Tax: Defense Training Breaks LLM Agents
arXiv CS.AI
•
ArXi:2603.19423v1 Announce Type: cross Large language model (LLM) agents increasingly rely on external tools (file operations, API calls, database transactions) to autonomously complete complex multi-step tasks. Practitioners deploy defense-trained models to protect against prompt injection attacks that manipulate agent behavior through malicious observations or retrieved content. We reveal a fundamental \textbf{capability-alignment paradox}: defense