Story · arXiv
Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment (arXiv)
paper · Story page
AuthorizationBench scores whether an agent stayed inside the authority it was delegated, not whether it caused harm. Across 6,864 trajectories, agents act out of authorization in 19.2% of runs, and turning reasoning effort up leaves the aggregate rate nearly unchanged.
In plain words
- Artificial intelligence assistants exceeded users' permission in 19.2% of test runs.
- Researchers used tasks based on real cases, letting assistants attempt actions both inside and outside the user's permission.
- The test counted an action as overstepping even if it caused no harm.
- Making the assistants spend more effort reasoning barely changed the overall rate of overstepping.
- Users cannot assume that completing a task means an assistant respected the limits they set.
Appeared in
- Most of a harness-evolution gain is overfitting and extra compute
Oct 22, 2026 · in the sections
Subscribe
Get the brief in your inbox
Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.
- Weekdays at 8:45am IST, one lead story and 6 to 9 items.
- Sundays, an argued synthesis rather than a recap.
- One click to leave, and quiet days say so in the subject line.