{
  "id": 6986,
  "url": "https://arxiv.org/abs/2603.19423v2",
  "title": "The Autonomy Tax: Defense Training Breaks LLM Agents",
  "summary": "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 f",
  "authors": "Shawn Li, Yue Zhao",
  "category": "research",
  "topics": "safety-alignment,military-security,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-03-19T19:33:17.000Z",
  "fetched_at": "2026-07-14T16:32:54.533Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/6986",
  "original_url": "https://arxiv.org/abs/2603.19423v2",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}