GLM-5.3 Uncovers 1,097 Critical Vulnerabilities Post-Training

Z.ai has released GLM-5.3, an open-weight AI model that unexpectedly identified 1,097 critical and high-severity vulnerabilities in real-world software. The model's capabilities were enhanced solely through post-training, leading to improved performance in cybersecurity tasks, including multi-step exploit-chain reasoning. This version has been reported to find vulnerabilities in widely used software like the Linux kernel and WinRAR, with some flaws dating back 45 years. The launch was delayed for a safety review due to the unexpected findings. While GLM-5.3 shows significant improvements over its predecessor, it still lags behind some competitors in specific benchmarks. The model is currently available only to paying customers, with a broader release planned after further security evaluations.

Key Points: • GLM-5.3 discovered 1,097 critical vulnerabilities in real-world software. • The model's enhancements came solely from post-training improvements. • Z.ai plans to release the model publicly after a two-week safety review.