China's Z.ai says new model nears Anthropic's Mythos 5 in cyber-defence tests

Beijing, Aug 14 - Chinese AI startup Z.ai claims its open-source GLM-5.3 model has come close to matching Anthropic's restricted Mythos 5 in detecting software vulnerabilities, according to a report on August 14. Z.ai stated that GLM-5.3 scored 84.5 percent on CyberGym, a test measuring the model's ability to review code, identify security flaws, and confirm their authenticity, slightly outperforming Mythos 5's 83.8 percent.

However, GLM-5.3 trailed behind Mythos 5 in transforming discovered flaws into functional attacks - a crucial component of defensive security research. Z.ai reported GLM-5.3 secured 54.4 percent on the ExploitBench test, compared to Mythos 5's 78.0 percent. In timed tests, GLM-5.3 handled 105 attack-development tasks within two hours and 130 in six, while Mythos 5 completed 181 and 247 tasks, respectively.

Z.ai plans to release GLM-5.3 publicly within two weeks after completing security assessments and fortifying its safety measures. The company added multiple security layers to GLM-5.3, including systems to screen risky requests, monitor the model's work, and train it to reject malicious tasks, aiming to distinguish harmful activity from legitimate uses.

Critics argue that safeguards may prove challenging to maintain once a model is released for anyone to download, modify, or combine with external tools. Z.ai insists that advanced cyber-defense tools should be accessible to developers of open-source software and smaller security teams, as opposed to being controlled by a few closed-model providers.