The BasePeople

Zhiping Zhang

Northeastern University

Titles as the sources gave them. In 1 entry, with 4 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“Agents act out of authorization in 19.2% of trajectories, with rates ranging from 10.5% to 26.6% across the 24 configurations.”
Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization AlignmentarXiv, October 20, 2026. Speaking as: Other. Headline result from the abstract: 6,864 trajectories from four models at two reasoning-effort settings on three agent harnesses.
“Increasing reasoning effort leaves the aggregate rate nearly unchanged (19.3% at low effort versus 19.1% at high effort), with effects varying across models.”
Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization AlignmentarXiv, October 20, 2026. Speaking as: Other. Abstract finding on whether more reasoning helps agents stay within granted authority.
“Authorization failures are therefore not confined to adversarial attacks.”
Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization AlignmentarXiv, October 20, 2026. Speaking as: Other. The authors' conclusion from finding that agents overstepped more often under benign pressure than under malicious injection.
“Agents that ask the principal before completing a task act out of authorization less often.”
Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization AlignmentarXiv, October 20, 2026. Speaking as: Other. The authors' sixth finding, reported as an association (14.0% versus 21.6% of runs), not a tested intervention.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    Act on Your Behalf, Beyond Your Authorization: Benchmarking LLM Agent Authorization Alignment

    Researchers at Northeastern, Carnegie Mellon and UNC-Chapel Hill introduce AuthorizationBench, 286 tasks built from 44 documented cases of AI agents overstepping instructions. Across 6,864 runs of four models, they report agents exceeded their granted scope in 19.2% of runs, whether or not harm followed.