GLM-5.3 Hype check
Anthropic says GLM-5.3's cyber safeguards are easy to bypass
50 of 410 exploit attempts Developing 2 sources
G
GLM-5.3
Model Sep 30, 2026
Z.ai's open-weight model assessed for cyber capabilities by NIST and Anthropic.
2 sources attached
What happened lately
GLM-5.3 timeline
What it is
Z.ai released GLM-5.3 on August 14 and published its weights roughly two weeks later, according to NIST’s Center for AI Standards and Innovation (CAISI). This entry tracks the model and its independently reported cyber evaluations; the September 29 Anthropic report is a separate news event, not a new model launch.
What the assessments say
CAISI called GLM-5.3 the most cyber-capable open-weight model it had evaluated, while finding it below the leading US frontier models in its benchmark comparison. Anthropic later reported strong exploit development results and tested ways to bypass or remove the model’s refusals. Those safeguard findings are Anthropic’s measurements, not an independent audit of every deployment.