Beijing-based AI startup Z.ai saw its stock drop by up to 9% in Hong Kong after announcing its new GLM-5.3 model. While the model shows strength in identifying software vulnerabilities, investor concerns regarding the safety and potential misuse of its upcoming public release appear to have weighed on the company’s share price.
Chinese AI firm Z.ai announced the launch of its GLM-5.3 model on Friday, claiming the software narrowly outperformed Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol in specific vulnerability-detection tests. According to the company's internal reports, GLM-5.3 achieved a score of 84.5% on the 'CyberGym' benchmark, slightly higher than the 83.8% reported for Mythos 5. Despite the technical milestone, the company’s stock saw a sharp reaction, falling by as much as 9% in Hong Kong trading during the session.
The market reaction highlights a growing tension between the rapid development of AI capabilities and the associated safety risks. While Z.ai's new model excels at finding software flaws, it trailed significantly behind competitors in the ability to convert those flaws into actionable cyberattacks. On the 'ExploitBench' test, which measures the capability to execute an attack, GLM-5.3 scored 54.4%, well below the 78.0% achieved by Anthropic's Mythos 5. This performance gap suggests that while the model is effective at identification, it currently lacks the advanced offensive utility of Western counterparts.
A key factor weighing on investor sentiment is Z.ai's intent to release the model's weights to the public in approximately two weeks, following a period of safety and security hardening. Unlike Anthropic's Mythos 5, which is a restricted tool available only to vetted organizations, Z.ai plans to offer wider access. Critics and investors have raised concerns that providing open access to such powerful code-auditing tools could inadvertently lower the barrier for malicious actors to develop automated cyber threats, regardless of the safeguards the company intends to implement.
From a financial perspective, the AI sector continues to face pressure due to the heavy capital spending required to train and maintain these large models. Investors are increasingly sensitive to companies that face regulatory scrutiny or potential reputational risks regarding AI safety. The 9% drop in Z.ai shares suggests that market sentiment is cautious about the potential fallout if the planned public release does not successfully address these security concerns.
The most important factor for investors to monitor in the coming weeks is the official public release scheduled for late August. Market observers will likely focus on whether the company’s 'trusted access' and safety protocols can genuinely prevent misuse, or if regulatory bodies will intervene. Additionally, any further updates on the model's performance in real-world scenarios versus the controlled CyberGym environment will be critical for assessing its long-term commercial value.