Skip to content
PandorIAThe box is already open
ENES
All entries

open weightsIn brief · · · Written with AI

Open AI models are six or seven months behind on cyberattack capabilities

An Epoch AI analysis puts the best open-weight model, GLM-5.3, where the best closed models were in the spring.

Epoch AI chart: record cybersecurity capability scores of AI models, January 2025 to October 2026. Closed models stay above open models throughout; the best open model, GLM-5.3, reaches about 160 points.
Image: Epoch AI (Jaeho Lee, Alexander Barry) · CC BY

According to an analysis by Epoch AI published on 16 October, the best open-weight model, Z.ai's GLM-5.3, is now where the best closed models were six to seven months ago on cybersecurity capabilities: above GPT-5.4, from March, and below Claude Mythos Preview, from April. The most capable closed models, Claude Opus 5.5 and GPT-6 Astra, remain ahead.

Anyone can download and modify an open-weight model. Epoch warns that GLM-5.3's score rests on just three test results, and that some of the data attributes to specific models results the evaluator published without naming them.

Sources

Written by AI from these sources.

  1. Open-weight models trail closed models in cyber capabilities by 6-7 monthsAccessed 17 October 2026 · Image: Epoch AI