Selectstar, an AI data and reliability assessment company, said on Wednesday its in-house red-teaming technology, Startiming, was officially adopted by ACL 2026, a top natural language processing (NLP) conference.

The accepted paper covers an automated red-teaming method to verify the safety of large language models (LLMs). Red-teaming is a safety evaluation method that attempts intentionally harmful requests to an AI model to find vulnerabilities.

Startiming uses statistical physics-based mathematical modelling to learn the relationship between attack strategies and model responses and probabilistically selects the optimal strategy. While existing approaches repeat past success cases, the company explained that this technology analyses numerous attempts and failures to find a strategy suited to the situation on its own.

In verification tests on 17 LLMs including Claude, Gemma, GPT, Llama and Qwen, Startiming achieved an average attack success rate of 74.5 percent based on a standard benchmark. That was 13.5 percentage points higher than the previous best method, AutoDAN-Turbo, at 61.0 percent.

The technology has been installed in Selectstar's AI reliability verification solution, the Datumo platform. It is also being applied to major domestic industries such as electronics and home appliance manufacturing, systems integration (SI) and IT services, as well as a government-led project for an independently developed AI foundation model.

Jung Min-jae (정민재), a Selectstar AI safety engineer and first author of the paper, said, "I wanted to present a structure that can find AI vulnerabilities more systematically." He added, "I will contribute to advancing the Datumo platform's technology so that LLMs can be used safely in real industrial settings."