Breaking

Aikido Security Releases Altar-1, an Open-Weight Cybersecurity Model for On-Prem Defense

September 22, 2026 at 07:04 EDT

Foundation Models
Security
AI Agents

Aikido Security has released Altar-1, an open-weight cybersecurity model designed to run defensive AI workloads inside customer-controlled infrastructure, including air-gapped environments. The model compresses Z.AI's GLM-5.3 from a 1.51 TB full-precision checkpoint to 328 GB while preserving much of the parent model's vulnerability-finding performance.

Sep. 21, 2026 · Aikido Security

Altar-1 shrinks local cyber AI

Aikido compressed a GLM-5.3-derived security model to 328 GB so defensive agents can run inside customer-controlled and air-gapped infrastructure.

328 GB

stored weights

78.2%

smaller than BF16

60.4%

average CVE recall

4x H200

serving target

Compression ladder

Stored weights fall from full GLM-5.3 to the Altar-1 prune.

1,506.7 GBGLM-5.3 BF16

488.2 GBAWQ INT4

328.0 GBAltar-1

How the model gets smaller

GLM-5.3 base

Large MoE parent model

→

AWQ INT4

Lower-bit stored weights

→

REAP prune

168 experts kept per layer

→

Local defense

Runs inside controlled networks

Why defenders care

Security agents can reason over source code, documentation and findings without sending sensitive context to an outside inference service.

What still limits it

The release is built for serious infrastructure, and the benchmark covers targeted CVE rediscovery rather than every step of real-world vulnerability discovery.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.