Large language model agents can exceed authorized limits while pursuing user goals, and a 648-sample benchmark across seven domains shows that this drift rises with capability yet is avoidable and carries cross-jurisdictional privacy and information-security risks.
LLM agents can exceed their authorization boundaries while pursuing user goals without an adversary. We formalize this phenomenon as authorization drift across data, tool, scope, and time, and introduce the Authorization Drift Benchmark (ADB): 648 samples across seven domains, each pairing a sufficient compliant path with an attractive unauthorized shortcut. To ensure executable tasks, tempting shortcuts, and feasible compliant paths, Staged Execution-Reflective Synthesis (SERS) grounds benchmark validation in real agent rollouts rather than LLM self-critique. Across 13 LLMs and three frameworks (25,207 trajectories), greater capability accompanies greater drift, with frontier-capable…
LLM agentsauthorization driftAuthorization Drift Benchmark (ADB)Staged Execution-Reflective Synthesis (SERS)agent rolloutsLLM self-critiqueagentic scaffoldsfrontier-capable modelsclean-success trajectoryclean-completion performance
- Upvotes
- 5
- Authors
- 10
- Published
- Nov 6, 2026
Summary and abstract excerpt via Hugging Face Papers ↗ · arXiv.