arXiv:2603.19423v2 Announce Type: replace-cross Abstract: 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 training designed to improve safety systematically destroys agent competence while failing to prevent sophisticated attacks. E
Source: arXiv cs.LG — read the full report at the original publisher.