Search results for CyberGym
AI & Enterprise
Security-specialised foundation model aims for Opus 4.5-class performance with Exaone, HyperCLOVA X
South Korea’s government-backed security-specialised AI foundation model will target global frontier-level performance, using Naver Cloud’s HyperCLOVA X for defence and LG AI Research’s Exaone for attack functions. The plan uses adversarial co-training in a feedback loop between the two models. The project will train on about 830TB of security data and use 4,512 GPUs from the government and private sector. It will be linked to security agents and tested across seven key sectors.
AI & Enterprise
Chinese open-weight models target security as Z.ai challenges Mythos with GLM-5.3
Chinese AI companies are accelerating efforts to expand their influence in the cybersecurity market by promoting open-weight models. Chinese startup Z.ai said on Aug. 15 it unveiled GLM-5.3, a cybersecurity-focused AI model, and plans to publish its weights within two weeks. The company said the model improved performance in vulnerability detection and exploit tasks, and set out a phased release plan with multi-layer safeguards and a separate disclosure process.
AI & Enterprise
OpenAI updates GPT-5.5-Cyber, expands collaboration with security firms
OpenAI updated its cybersecurity model GPT-5.5-Cyber and expanded how vetted security companies can use it, Axios reported on June 22. OpenAI said the update lets approved security professionals do more work and improves performance in security tasks. The new version supports deep analysis of large codebases, vulnerability identification and verification, and patch development and testing. OpenAI also introduced a partner program and backed an initiative with Trail of Bits, HackerOne and Calypso.
