{
  "id": 12362,
  "url": "https://arxiv.org/abs/2607.18366v1",
  "title": "Operational Hallucination and Safety Drift in AI Agents",
  "summary": "Large language models (LLMs) serving as planners in tool-using autonomous agents introduce dynamic reliability risks in multi-turn execution. While single-turn safety mechanisms are relatively mature, extended interactions reveal structural vulnerabilities where initial alignment degrades over time. This paper empirically characterizes two observed failure modes across multiple state-of-the-art LLMs: Safety Drift, the gradual erosion of declared safety intent leading to constraint-violating acti",
  "authors": "Shasha Yu, Fiona Carroll, Barry L. Bentley",
  "category": "research",
  "topics": "safety-alignment,agents-autonomy",
  "orgs": null,
  "regions": null,
  "published_at": "2026-07-20T17:01:50.000Z",
  "fetched_at": "2026-07-22T05:10:49.469Z",
  "source_slug": "arxiv-ethics",
  "source_name": "arXiv",
  "source_homepage": "https://arxiv.org",
  "ethics_ai_record_url": "https://ethics.ai/record/12362",
  "original_url": "https://arxiv.org/abs/2607.18366v1",
  "evidence_status": "source-only",
  "attribution": "via ethics.ai"
}