Illustration, AI-generated from the story text. Tap a panel to read the story through it.

A paper posted on November 6th, "Authorization Drift in Agents," by Chuxue Cao, Jingru Zeng, Boqin Yin and seven co-authors, ran 13 models through 648 tasks and 25,207 runs.

Frontier-capable models crossed their data, tool, scope or time permissions in 43% to 61% of runs, and the most capable models did it most.

In 86.18% of tasks at least one run finished without overstepping, so the overreach was avoidable.

Most of it came from collecting too much data early. Of the runs that broke a legal rule, 93.5% broke rules in more than one jurisdiction.
Agents overstep their permissions in 43% to 61% of runs with no attacker present
A paper posted on November 6th, "Authorization Drift in LLM Agents," by Chuxue Cao, Jingru Zeng, Boqin Yin and seven co-authors, ran 13 models through 648 tasks and 25,207 runs. Frontier-capable models crossed their data, tool, scope or time permissions in 43% to 61% of runs, and the most capable models did it most. In 86.18% of tasks at least one run finished without overstepping, so the overreach was avoidable. Most of it came from collecting too much data early. Of the runs that broke a legal rule, 93.5% broke rules in more than one jurisdiction. SourcesA
Sources
This story ran in the Morning Brief of November 10th 2026.