Open Models Close Cyber Gap As GLM-5.3 Hits 160

Zai‘s GLM-5.3 shows Open Models trailing the closed frontier in cybersecurity capability by roughly six to seven months, according to Epoch AI Research.

Key Takeaways

Zai’s GLM-5.3 reached 160 on Epoch AI Research’s Cyber ECI, the highest score among open-weight models
Epoch AI Research estimated open models trail the closed frontier in cybersecurity capability by roughly six to seven months
Open-weight models publish underlying parameters publicly, allowing anyone to download, inspect or modify them
A narrowing open-model gap shifts defensive burden toward network-level and endpoint security rather than AI-provider-level controls

Zai’s GLM-5.3 posted the highest score of any open-weight model on the test, reaching 160 on a cyber capability index built from closed-model baselines. Epoch AI Research, an organization that tracks AI progress through independent benchmarking, posted the finding on X on Oct. 16.

Open-weight models are AI systems whose underlying parameters are published publicly, letting anyone download, inspect or modify them, in contrast to closed models from firms like OpenAI that run only through a company’s own servers and API.

The group’s Cyber ECI measures how capable a model is at offensive and defensive cybersecurity tasks, including vulnerability discovery and exploit development, scored against a baseline tied to closed-frontier performance over time.

A six-to-seven month gap is narrow by historical standards for open-weight AI. Through most of 2024 and early 2025, open models trailed closed frontier systems in specialized domains like coding and reasoning by closer to a year or more.

The compression on cyber tasks specifically raises the stakes because offensive cyber capability has direct applications for both defenders and attackers.

A shrinking gap means powerful cyber tooling is available to anyone who can run an open-weight model rather than only to customers of frontier labs.

Why Open Models’ Shrinking Cyber Capability Gap Changes The Threat Model

Security researchers and AI labs have spent the past year warning about AI-assisted hacking as closed models grew capable enough to automate parts of vulnerability research and exploit writing.

Also Read: Liquid AI Bets On Privacy With 24B On-Device Model, Skips Benchmark Claims

OpenAI and others have built internal teams to monitor misuse of their own frontier models for offensive cyber purposes, treating the closed-model gap as a kind of safety buffer. Anyone attempting large-scale misuse previously needed access to the most capable systems, which labs can monitor and restrict.

A narrowing gap on open-weight models erodes that buffer.

Once a capability is available in a model anyone can download and run locally, no single company can monitor, rate-limit or revoke access to it.

That shifts the defensive burden toward network-level and endpoint security rather than AI-provider-level controls, a distinction cybersecurity teams have flagged as Open Models have grown more capable across coding, reasoning and cyber-specific tasks over the past 18 months.

A report from security firm ClawSecure published separately this week found flagship-class agents from multiple labs capable of autonomously hiding intrusion activity, a related but distinct finding on agent behavior rather than raw capability scoring.

How Zai’s Open Models Closed Ground

Zai has released successive GLM model versions over the past year aimed at narrowing gaps with closed-frontier systems across coding and reasoning benchmarks, and the GLM-5.3 result extends that trajectory into cybersecurity.

The company has positioned its open releases as a direct competitor to Chinese and Western closed labs alike, part of a broader wave of open-weight Chinese model releases that have compressed capability gaps across nearly every benchmark category tracked by independent evaluators through 2026.

What Could Widen Or Close The Gap Next

Epoch AI’s index will update as new frontier and open releases land, and the next test is whether an upcoming closed-lab release widens the gap back out or another open model closes it further. Policymakers debating export controls and model-weight restrictions will likely treat benchmarks like the Cyber ECI as evidence in that fight.