- Z.ai announced that its open-source GLM-5.3 model, released on August 14, has neared Anthropic's restricted Mythos 5 in identifying software vulnerabilities.
- Z.ai reported that GLM-5.3 scored 84.5% on CyberGym, slightly surpassing Mythos 5's 83.8%, but lagged behind on ExploitBench with 54.4% compared to 78.0% for Mythos 5.
- The company plans to publicly release GLM-5.3 in about two weeks after completing security assessments, with sensitive functions available only through a 'trusted access' program.
- Z.ai framed the launch of GLM-5.3 as a challenge to restricted-access models, introducing an 'Open Source Shield' initiative to support open-source software developers.
- GLM-5.3 is based on the same model as GLM-5.2 but has improved its programming capabilities by 50% through expanded post-training and reinforcement learning.
- Anthropic has restricted access to Mythos, a version of its Claude Fable 5 model, to vetted organizations due to concerns about the potential misuse of AI systems capable of finding and exploiting software flaws.
- Z.ai has added several layers of protection to GLM-5.3, including systems to screen risky requests and monitor the model's work to distinguish harmful activity from legitimate uses.
Chinese AI startup Z.ai has made significant strides with its open-source GLM-5.3 model, claiming it is close to matching Anthropic's Mythos 5 in software vulnerability detection. The model achieved an 84.5% score on CyberGym, surpassing Mythos 5's 83.8% score, although the results remain unverified.12345678
However, GLM-5.3 fell short in converting identified flaws into actionable attacks, scoring 54.4% on ExploitBench compared to Mythos 5's 78.0%. In timed tests, GLM-5.3 completed 105 attack-development tasks in two hours and 130 in six hours, while Mythos 5 completed 181 and 247 tasks, respectively.
Z.ai plans to publicly release GLM-5.3 in two weeks, following security assessments and enhancements to its safeguards. The company has implemented several protective measures, including systems to screen risky requests and monitor the model's outputs, aimed at distinguishing harmful activities from legitimate uses.
Z.ai positions this launch as a challenge to the restricted access of Mythos, advocating for the availability of advanced cyber-defense tools to open-source developers and smaller security teams. The company emphasizes that such tools should not be monopolized by a few closed-model providers, promoting a more inclusive approach to cybersecurity.
“GLM-5.3 scored 84.5% on CyberGym, edging Mythos 5's 83.8%, but lagged on ExploitBench (54.4% vs 78.0%). Z.ai will release weights publicly after security checks, with sensitive functions gated behind a 'trusted access' program.”
CNBC TV18
