GLM-5.3 Boosts Coding Agents, Flags Critical Cursor Vulnerability

GLM-5.3 Raises the Bar for Coding Agents

Z.ai’s latest release, GLM-5.3, delivers a dramatic leap in long‑horizon coding and unexpected strength in cybersecurity, spotting a potentially serious vulnerability in SpaceX‑acquired AI editor Cursor, according to developer advocate Lou on X. Reuters has reached out for comment.

GLM-5.3 Cybersecurity Leap Finds Cursor Flaw

Post‑training scaling, not a fresh pre‑training run, fuels the boost: the base 743‑billion‑parameter model of GLM‑5.2 remains unchanged, while reinforcement‑learning across diverse codebases expands its reasoning horizon. On the internal CyberGym benchmark the model scores ↑ 84.5%, edging out the prior ↓ 77.2% and surpassing rivals such as GPT‑5.6 Sol.

“As we scaled post‑training, cyber capability developed faster than we expected,” Z.ai noted.

Enterprise access starts via the GLM Coding Plan and ZCode environment; API keys and open weights are slated for release after a safety hardening phase, with weights expected pandemic‑adjacent timeline of roughly two weeks.

Pricing tiers begin at $12.60 per month for 10,000 weekly credits, with higher plans offering up to 14× usage. Token consumption improves markedly: the Max reasoning mode now uses ~75,000 output tokens per task versus 96,000 in GLM‑5.2, cutting inference costs for long‑running agents.

Words by: Dr. Aris Thorne

Artificial Intelligence Researcher