ai_security▌

← feed

NOTABLE

arXiv: Authorization Drift Benchmark measures LLM agents crossing permission boundaries

arXiv 2611.09113 formalizes authorization drift — agents exceeding their authorization across data, tool, scope and time while pursuing user goals — and benchmarks 648 samples across seven domains pairing compliant paths with attractive unauthorized shortcuts.

released 2026-11-06 · ingested 2026-11-09T12:17Z

Sources

Primary source

Cite this

APA

AI Security Index. (2026). arXiv: Authorization Drift Benchmark measures LLM agents crossing permission boundaries. AI Security Index. https://arxiv.org/abs/2611.09113

BibTeX

@misc{aisec-2026-11-09-arxiv-2611-09113-authorization-drift-llm-agents-benchmark,
  title  = {arXiv: Authorization Drift Benchmark measures LLM agents crossing permission boundaries},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2611.09113}},
  note   = {AI Security Index item 2026-11-09-arxiv-2611-09113-authorization-drift-llm-agents-benchmark}
}