FULL STORY
GLM-5.3 Launch and the Open-Source Shield Initiative
Zhipu released GLM-5.3 alongside an Open-Source Shield safety program responding to the Dragonfly plan. Cybersecurity benchmarks soon surfaced, prompting Zhipu to restrict access to the model's offensive security capabilities.
2026-08-18 ~ 2026-08-19 · 2 episodes · 5 posts
Episode 1 · Z.ai Launches GLM-5.3 with Open-Source Shield Safety Program (2026-08-18, 3 posts)
Z.ai launched GLM-5.3 along with the 'Open-Source Shield' program, offering free security audits, automated code auditing tools, and model credits to the open-source community. The model features a tiered risk review system that blocks high-risk requests while restricting sensitive offensive capabilities to verified users.
- Zhipu AI Launches 'Shield of Open Source' with Free Audits and Quotas — pstAsiatech · 2026-08-18
- Zhipu Launches 'Shield of Open Source' to Counter Cyber Risks — pstAsiatech · 2026-08-19
- Zhipu adds layered risk review to GLM-5.3, acting on AI safety without government push — pstAsiatech · 2026-08-19
Episode 2 · GLM-5.3 Cyber Eval Released as Z.ai Restricts Offensive Capabilities (2026-08-19, 2 posts)
GLM-5.3 scored 84.5% on CyberGym, slightly above Mythos 5's reported 83.8%, though it lags in exploit development. Z.ai is restricting its most sensitive offensive cyber capabilities to verified users under the 'Cybersecurity Trusted Access' program.
- GLM-5.3 matches Mythos on flaw detection but trails badly on weaponizing exploits — pstAsiatech · 2026-08-19
- Zhipu's "Cybersecurity Trusted Access" restricts offensive capabilities to verified users — pstAsiatech · 2026-08-19