Zhipu launches flagship model GLM-5.3 as China seeks Mythos-level edge in cyber defence

Chinese AI firm Zhipu, also known as Z.ai, has unveiled its flagship model GLM-5.3, claiming it outperformed Anthropic's Mythos 5 in a crucial cybersecurity test. The company's claim is based on GLM-5.3's 84.5% success rate on CyberGym, a benchmark that assesses models' ability to detect and validate security flaws in source code.

This surpassed both Mythos 5's 83.8% and OpenAI's GPT-5.6 Sol's 83.6%. However, GLM-5.3 lagged behind Mythos 5 and GPT-5.6 Sol in ExploitBench, which measures how far AI models can exploit vulnerabilities, with scores of 54.4% and 76.5% respectively.

Zhipu claims that GLM-5.3 was tested with Chinese security teams against real-world codebases, identifying 2,436 vulnerabilities across 269 projects, with 1,097 rated as medium to high severity. The company plans to release GLM-5.3's weights in about two weeks following an internal safety review and hardening process. Zhipu describes GLM-5.3's cyber-defensive capabilities as "public goods" open-source for worldwide developers to enhance.

Zhipu emphasizes that AI development should be a collaborative effort among nations rather than a solitary pursuit. The company hopes that making GLM-5.3 open-source will prevent it from being a privilege of a select few. GLM-5.3 shares the same base model as its predecessor GLM-5.2, a 744-billion-parameter model introduced in June.

GLM-5.3's coding abilities have approached those of Anthropic's Fable 5, according to Zhipu. The model scored 28.5 on the Agent's Last Exam, becoming the best open-weight model, ahead of Kimi K3's 27.6. GLM-5.3 narrowly trailed GPT-5.6-Sol, which scored 28.6.