🔒Z.ai Launches GLM-5.3 With Cyber Capabilities and Coding Gains
New Model Finds Vulnerabilities in SpaceX's Cursor AI
TL;DR
Z.ai launches GLM-5.3 with significant improvements in cybersecurity and coding efficiency. The model finds vulnerabilities in SpaceX’s Cursor AI and boosts coding performance benchmarks.
Z.ai has released GLM-5.3, a major update to its AI coding platform that significantly enhances both cybersecurity capabilities and long-horizon coding tasks. This new version identifies critical vulnerabilities in Cursor, an AI coding startup acquired by SpaceX, showcasing the model's advanced security features. For developers, this means improved efficiency and accuracy in code generation, with notable gains on Terminal-Bench 3.0 (jumping from 4.6 to 28.3) and DeepSWE v1.1 (from 46.2 to 66.9). The model also excels at cybersecurity tasks, scoring 54.4% on ExploitBench compared to GLM-5.2's 24.4%. However, developers need to adjust their API calls for the new version.
Key Points
GLM-5.3 identifies 2,436 vulnerabilities across 269 projects after expert review, with 1,097 classified as critical or high severity.
The new version scores 84.5% on CyberGym and 54.4% on ExploitBench, more than doubling GLM-5.2's performance in cybersecurity tasks.
GLM-5.3 boosts coding benchmarks: Terminal-Bench 3.0 score jumps from 4.6 to 28.3; DeepSWE v1.1 improves from 46.2 to 66.9.
Developers must adjust API calls for GLM-5.3, changing thinking.type from 'disabled' to 'enabled' and specifying reasoning effort levels.
GLM-5.3 supports three reasoning-effort levels: low, high (default), and max; max is recommended for coding tasks.
Why It Matters
If you're using Z.ai's GLM Coding Plan or the ZCode environment, GLM-5.3 offers substantial improvements in both cybersecurity and coding efficiency. For instance, on Terminal-Bench 3.0, GLM-5.3 scores a significant 28.3 compared to 4.6 for its predecessor. However, migrating existing applications requires adjusting API calls due to breaking changes.
Frequently Asked Questions
Why does this matter?
If you're using Z.ai's GLM Coding Plan or the ZCode environment, GLM-5.3 offers substantial improvements in both cybersecurity and coding efficiency. For instance, on Terminal-Bench 3.0, GLM-5.3 scores a significant 28.3 compared to 4.6 for its predecessor. However, migrating existing applications requires adjusting API calls due to breaking changes.
What happened?
Z.ai launches GLM-5.3 with significant improvements in cybersecurity and coding efficiency. The model finds vulnerabilities in SpaceX’s Cursor AI and boosts coding performance benchmarks.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,009 builders reading daily.