The authors constructed AuthGateBench, a benchmark of 1,400 instances spanning five failure modes drawn from real-world Claude Code incidents.
LLM authorization gates frequently block benign actions when safety alignment overrides explicit user permission
AuthGateBench tests classifiers on 1,400 instances across five failure modes, finding that models reject even explicitly authorized actions due to reliance on internal safety training rather than policy reasoning.
Chinese Tech
Zhixiong Zhuang · Jin Jia · Zhiwei Shang · Ziyu Wang
Huawei Technologies Ltd. · Sun Yat-sen University
Research Digest··2 min read
Thread:Agent Security & Attacks
The authors present AuthGateBench, the first policy-guided benchmark for evaluating LLMs as authorization gates in coding agents.
Why this paper
From Huawei Technologies Ltd. and Sun Yat-sen University · Part of Agent Security & Attacks, now 86 papers
In one line
Authorization gates using LLMs for coding agents often block explicitly authorized actions due to over-reliance on internal safety alignment, and retrospective environment signals improve accuracy.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§