Operational Hallucination and Safety Drift in AI Agents

Abstract

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

Categories

@misc{yu2026operational,
  title = {{Operational Hallucination and Safety Drift in AI Agents}},
  author = {Shasha Yu and Fiona Carroll and Barry L. Bentley},
  year = {2026},
  month = jul,
  eprint = {2607.18366},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.18366}
}