Open-weight AI models are 6-7 months behind on cyber
Published 24 October 2026
Z.ai's GLM-5.3 leads the open-weight Cyber ECI, reaching a score that closed frontier models achieved roughly six to seven months earlier.
Read the original on epoch.ai →
Boring AI News
More from 24 October 2026
- Introducing Flint
- Muse Glimmer 30B: An On-Device Agentic Model
- Four evaluation engineering lessons learned from building agent plugins
- Beyond Top-K: Building search-confidence guardrails for agentic AI with OpenSearch
- RAG Evals: Change Prompts Without Breaking Your App
- Cutting our inference bill by 68% by being obsessive about optimization
- Reducing Reward Hacking by Localizing Bad Behavior Instead of Penalizing It
- Security Audit Skill Enables Independently Verified Findings
- Plan or Retrieve First? A Scale Study of Tool Discovery Architectures
- Mistral's new forecasting models run on CPUs, with sizes starting at 5M parameters
- Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming
Every Monday we pick the five stories from the week worth keeping.
Read the latest AI Digest