- Published on
GLM-5.3 is here with advanced cyber capabilities — and reportedly already found a 'serious vulnerability' in Cursor
- Authors
- Name
- DeepHorus
Chinese AI startup Z.ai today released GLM-5.3, a substantial update to its open-source GLM series that delivers major gains in long-horizon coding and, more strikingly, advanced cybersecurity capabilities. The company says the model has already identified a “potentially serious vulnerability” in Cursor, the AI coding assistant recently acquired by SpaceX, according to Z.ai developer advocate Lou on X. VentureBeat has reached out to Cursor for comment and is awaiting a response.
GLM-5.3 is initially available only through Z.ai’s GLM Coding Plan and ZCode environment, with API access and open weights to follow once safety evaluation and hardening are complete. The company plans to release weights roughly two weeks after launch. For enterprise developers, the bigger story is not just another benchmark bump but how the gains were achieved: Z.ai kept the same 743-billion-parameter base model as GLM-5.2 and improved performance entirely through scaled post-training, including more diverse tasks and additional reinforcement-learning compute.
That approach positions GLM-5.3 as a test of how far a frontier-scale base model can be pushed without an expensive pretraining cycle. “Scaling post-training is all we did for GLM-5.3,” Z.ai wrote in its technical announcement. The results show considerable headroom, but they also created an unusual dilemma for an open-model developer: cybersecurity capabilities improved faster than expected, especially as tasks moved from vulnerability identification to building full exploitation chains. Reuters reported Friday that Z.ai is adding controls around these advanced capabilities, including a “trusted access” approach for sensitive functionality.
On benchmark evaluations, GLM-5.3 improved from 4.6 to 28.3 on Terminal-Bench 3.0, from 46.2 to 66.9 on DeepSWE v1.1, and from 26.2 to 48.2 on AutomationBench. It still trails some competitors, with GPT-5.6 Sol scoring 34.6 and Claude Fable 5 scoring 33.7 on Terminal-Bench. But Z.ai also emphasizes efficiency: on its private Code Bench, GLM-5.3 hits 34.5% at Max reasoning while consuming about 75,000 output tokens per task, versus GLM-5.2’s 23.4% using 96,000 tokens.
These company-reported figures should be treated with caution, but lower token consumption with better results matters practically for enterprises deploying coding agents at scale. The post-training environments now resemble complete engineering jobs, with agents accessing codebases, documentation, compute clusters, and experimental results, then diagnosing issues and demonstrating measurable improvements. Industry reaction is mixed: some praise the efficiency gains and rapid iteration, while others worry about the security implications of releasing powerful cyber tools in open weights.
Looking ahead, Z.ai’s two-week countdown to open weights will be watched closely by both enterprise adopters and security researchers. The release also underscores a broader trend: scaling post-training rather than base models is becoming a powerful lever for capability gains, but it magnifies safety questions that open-source developers must address. Cursor’s response, when it comes, will likely add another chapter to this fast-moving story.
