Security
Chinese AI Model GLM-5.3 Rivals U.S. Systems in Cybersecurity Tasks
Z.ai delays public release of open-weight model that scores 84.5% on vulnerability detection benchmark, citing security concerns.
Key takeaways
- Z.ai's GLM-5.3 scored 84.5% on vulnerability detection benchmarks, surpassing Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol.
- The Chinese lab is delaying public release by two weeks to implement safety controls, but acknowledges it cannot control the model once weights are public.
- Z.ai claims its GLM models have found over 2,400 security flaws, including more than 1,000 critical or high-severity vulnerabilities.
- Chinese labs are advancing open-weight cyber-capable models while some U.S. companies slow releases over security concerns.
- The Trump administration is reportedly considering regulations for open-source AI models as capabilities approach frontier systems.
A Chinese AI laboratory has developed a model with cybersecurity capabilities approaching those of leading U.S. systems, raising questions about the global race to control dual-use AI technology.
Z.ai announced Friday that its latest model, GLM-5.3, will face a two-week delay before public release while the company implements additional safety controls. The decision follows internal testing that revealed the model's advanced ability to identify and exploit security vulnerabilities.
Why it matters
Cyber-capable AI models represent a double-edged technology that can accelerate both defensive security operations and malicious hacking at scale. As Chinese labs advance open-weight models with fewer release restrictions than U.S. counterparts, the window for controlling access to these capabilities may be narrowing.
Benchmark performance
GLM-5.3 achieved an 84.5% score on CyberGym, a benchmark measuring how effectively models detect known security vulnerabilities. That performance surpassed both Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol.
On ExploitBench, which evaluates models' ability to reason through and develop exploits for real vulnerabilities, GLM-5.3 ranked third among tested systems, trailing only the two U.S. frontier models.
Z.ai specifically trained the model for vulnerability detection by allowing it to practice cyber tasks in controlled environments. The company is implementing tiered access that initially limits GLM-5.3 to selected security partners in controlled settings.
Real-world impact
According to a disclosure site released Friday, Z.ai claims its GLM models have identified more than 2,400 security flaws, including over 1,000 classified as critical or high severity. These discoveries span the Linux kernel and widely deployed VMware and Apache projects.
The company announced a new program allowing open-source maintainers to scan their repositories for bugs using GLM models.
Z.ai has positioned the release as beneficial for defenders, stating on X that "an open world cannot have only open attack surfaces" but "must also have an open shield."
The control problem
Once GLM-5.3's weights become publicly available, Z.ai acknowledged it cannot control how users modify or deploy the model. This limitation highlights a fundamental tension in open-weight AI development: maximizing beneficial access while preventing malicious use.
The timing is notable as some U.S. laboratories slow model releases over security concerns, while Chinese labs continue advancing open-weight models optimized for cyber tasks. Earlier this week, researchers at Dream reported that open-source AI agents were used in an automated cyberattack against Taiwan's government.
At least one GLM model has already proven useful in defensive operations. Hugging Face reported using GLM-5.2 to investigate a recent breach involving OpenAI's models after guardrails on U.S. frontier models declined to assist.
The Trump administration is reportedly considering regulatory approaches for open-source models as their capabilities increasingly rival closed systems.
These details were first reported by Axios.
