Chinese AI startup Z.ai announced that its latest open-source model, GLM-5.3, nearly matches the cybersecurity performance of Anthropic's restricted Mythos 5 in vulnerability detection benchmarks, signaling a new wave of competition in AI-powered cyber defence.
- GLM-5.3 scored 84.5% on vulnerability detection vs. Mythos 5’s 83.8%
- Z.ai to limit sensitive features to verified users for security reasons
- Open-source approach challenges restricted access models like Mythos
What happened
Chinese startup Z.ai reported that its newly developed GLM-5.3 AI model performed comparably to Anthropic’s Mythos 5 in cybersecurity evaluations. The GLM-5.3 scored 84.5% in detecting vulnerabilities on CyberGym, slightly edging out Mythos 5’s 83.8%. However, in tests focused on exploiting detected vulnerabilities, GLM-5.3 lagged behind, with a lower success rate in turning vulnerabilities into working attacks.
Z.ai also highlighted timed attack-development tests, where Mythos 5 completed substantially more tasks than GLM-5.3 within two and six hours. Despite this, Z.ai is preparing to release GLM-5.3 publicly after adding multiple security layers and plans to restrict access to its most sensitive cybersecurity functions via a trusted user program.
Why it matters
The development of GLM-5.3 challenges the dominance of closed-source cyber-defence models like Anthropic’s Mythos, which is available only to vetted organizations due to the dual-use risk of AI tools that can both identify and exploit software flaws. Z.ai’s move suggests a shift toward more open and accessible advanced AI cybersecurity tools, potentially empowering developers of open-source projects and smaller security teams.
Z.ai’s approach to delaying model release for safety assessments and implementing safeguards illustrates increasing sophistication in China's risk management practices for powerful AI models. This marks a significant step in AI governance, demonstrating that Chinese labs are considering safety similarly to their Western counterparts.
What to watch next
The upcoming public release of GLM-5.3 will be a key event, particularly how effectively Z.ai enforces its safeguards and controls once the model is available widely. Observers will watch if trusted access programs and safety monitoring can prevent misuse without stifling legitimate defensive applications.
Additionally, the broader cybersecurity community will follow how this open-source approach influences the balance of power between proprietary and open models in cyber-defence, as well as how governments and enterprises respond to access to sophisticated AI tools capable of both protecting and threatening software security.