Aikido Security introduces Altar, an open-weight security model that runs within customers’ own infrastructure. The model is a compressed version of GLM-5.3, reducing the size from 1.51 TB to 328 GB while retaining 92 percent of the vulnerability coverage of the original model.
The Ghent-based security company deploys Altar on its autonomous penetration testing appliance, the Aikido Machine, which runs within the customer’s infrastructure and also operates in fully air-gapped environments. This ensures that source code, architecture documentation, and unpatched findings never leave the customer’s own network.
That is precisely the bottleneck that Aikido aims to eliminate. Closed-frontier models, by their very nature, run on third-party infrastructure. For banks with data residency requirements, hospital groups, or industrial OT environments without an internet connection, that is simply not an option.
From 1.51 TB to 328 GB
Aikido started with Z.AI’s GLM-5.3, which scores 84.5 percent on CyberGym and 54.4 percent on ExploitBench. Quantization reduced the model to 488 GB, after which expert pruning via Cerebras REAP shaved off another 160 GB. Altar retains 168 of the original 256 routed experts per layer.
To determine which experts to remove, Aikido used traces from its own penetration testing harness, supplemented with multilingual text so the model can continue to understand documentation in languages such as French and Dutch. No customer data was used.
One percentage point lower recall
On an internal CVE benchmark with 32 vulnerabilities across 30 repositories, Altar achieved an average recall of 60.4 percent and identified 23 vulnerabilities. The quantized GLM-5.3 achieved 61.5 percent with the same coverage, while the full model reached 65.6 percent and identified 25 vulnerabilities.
Following the evaluations, Aikido deployed Altar across its Aikido Machine fleet. Shortly thereafter, the model identified a valid vulnerability with critical severity during a production penetration test at a client site.
Altar’s weights are managed by Aikido. According to the company, the model runs on a node with four H200s and the latest version of vLLM. Aikido Labs is further investigating lower-bit formats such as EXL3 and fine-tuning for security workflows.
Tip: Aikido Endpoint offers developers additional protection against supply chain attacks