Korea Advanced Institute of Science and Technology said it developed a safety verification technology that can uncover about seven times more diverse hidden vulnerabilities in generative AI than existing methods. The team built a new red-teaming framework, Stable-GFlowNet, to verify large language model safety. Tests found 134 unique attack types versus 17 previously, while maintaining a 92 percent attack success rate. The work was selected as an ICML 2026 spotlight paper.

KAIST said a research team developed a learning framework designed to improve the safety of customised AI trained on personal or corporate documents while keeping task performance. The Buffer-and-Reinforce approach applies a temporary buffering module during fine-tuning to block malicious data from affecting the core model, then removes it and adds a safety reinforcement module using QR decomposition. In tests, the share of dangerous answers fell to about 8 percent from about 18 percent.