AuthGateBench, described in a new arXiv paper, targets a specific security problem: deciding whether a coding agent should be allowed to perform a high-risk action. The authors frame this as an authorization gate, where a classifier—typically a large language model—replaces or reduces the need for human approval. Such systems are already adopted in developer tools, according to the abstract.

The benchmark appears designed to measure how reliably LLMs can act as these gates. Because the decision is delegated to the model, errors could let unsafe actions through or block legitimate work. The paper's contribution is a standardized way to evaluate that trade-off, though the abstract does not report specific results.