Aikido Releases Altar, a Local Cybersecurity AI Model Built From GLM-5.3

Aikido’s promise with Altar is straightforward: use AI to inspect security-sensitive code without handing that code to an outside inference provider. The newly released open-weight model is designed for local deployment, but the evidence for how well its compression holds up is limited to Aikido’s own targeted vulnerability benchmark.

A model sized for the security workflow

The Belgian cybersecurity company released Altar on September 21 as its first open-weight security model. It is a compressed and customized version of Z.AI’s open-weight GLM-5.3, intended for work such as reviewing code for vulnerabilities and supporting penetration-testing workflows inside infrastructure a customer controls.

Aikido says it combined quantization, which stores model weights with less precision, with expert pruning, which removes selected specialized parts of a mixture-of-experts model. Altar retains 168 of the parent model’s 256 routed experts in each backbone expert layer; the company says the selection was calibrated with internal pentesting traces and multilingual text, not customer data.

A narrow test of what survived compression

Aikido evaluated Altar on an internal benchmark of 32 known vulnerabilities across 30 code repositories. The company says Altar averaged 60.4% recall per run and rediscovered 23 vulnerabilities at least once across three runs. Its quantized GLM-5.3 baseline averaged 61.5% recall and covered the same 23 vulnerabilities; the full-precision parent averaged 65.6% recall and covered 25.

What Aikido’s benchmark does not measure

• Blind vulnerability discovery across an entire codebase.
• Whether the model can execute exploits to validate a finding.
• The quality of the fix proposals that follow a finding.

Those boundaries matter because the results are a measure of rediscovering known flaws within a pipeline that uses other models for surrounding stages. They support Aikido’s claim that much of the parent model’s tested vulnerability coverage remained after compression, not a broad demonstration of end-to-end autonomous security performance.

The deployment constraint is part of the pitch

The release targets organizations that cannot, or do not want to, send source code, internal architecture documents or unresolved findings to a third-party AI service. Aikido says Altar can be served on a node with four H200 GPUs, and that its weights are publicly available under the company’s organization.

Aikido plans to deploy Altar in its own products, including tools used by customers such as Belgian bank Belfius. The practical test is now whether a smaller model running on customer-controlled hardware can deliver sufficiently reliable security findings to justify the added operational burden of hosting it.