Aikido Security Unveils Altar-1 Open-Weight AI for Cybersecurity Defense

Aikido Security has unveiled Altar-1, an open-weight artificial intelligence model designed to run defensive cybersecurity workloads entirely inside an organization’s own infrastructure.

The model aims to help security teams use advanced AI for vulnerability discovery and penetration testing without sending source code, internal documentation, or security findings to external cloud-based inference services.

Altar-1 is built for organizations with strict privacy, regulatory, or operational requirements. These include banks with data-residency obligations, healthcare providers handling sensitive patient information, and industrial organizations with isolated or air-gapped operational technology environments.

By keeping both the model and the data under customer control, Aikido is positioning Altar-1 as a sovereign security intelligence option.
The model powers Aikido Machine, the company’s autonomous pentesting appliance.

Aikido Machine is designed to operate within customer environments and continuously identify, exploit, and validate weaknesses across an attack surface.

Altar-1 adds local AI reasoning capabilities to that workflow, allowing security testing to process sensitive technical context without transmitting it to a third party.

Aikido Security Unveils Altar-1

Aikido created Altar-1 from Z.AI’s open-weight GLM-5.3 model, reducing its storage footprint from 1.51 TB to 328 GB. The company first applied AWQ INT4 quantization, reducing the model to 488.2 GB, then used expert pruning to remove components less relevant to targeted security workloads.

Mixture-of-experts models contain many specialist neural-network components, known as experts. During inference, only a limited number are selected for each token. However, the larger expert pool normally still requires substantial memory.

Aikido said it retained 168 of the original 256 routed experts in each backbone layer and removed 88 experts. The result represents a 78.2 percent reduction from the original full-precision model and a 32.8 percent reduction from the already quantized checkpoint.

To decide which experts to preserve, Aikido used traces from internal pentesting benchmarks rather than customer data. These traces included representative code, tool calls, and agent responses generated during security testing workflows.

The company also incorporated multilingual text during calibration to retain the language understanding needed to review documentation, business rules, application behavior, and user interfaces in different languages.

Aikido used Cerebras REAP (Router-weighted Expert Activation Pruning) to select experts based on router weights and output magnitude, preserving coding, cybersecurity, and natural-language reasoning capabilities.

In an internal benchmark covering 32 known vulnerabilities across 30 repositories, Altar-1 achieved an average recall of 60.4 percent per run. It rediscovered 23 of the 32 vulnerabilities at least once across three runs.

The quantized but unpruned GLM-5.3 model reached 61.5 percent recall and found the same 23 vulnerabilities. In comparison, the full-precision parent model achieved 65.6 percent recall and identified 25 vulnerabilities.

Aikido said Altar-1 therefore retained 92 percent of the full model’s vulnerability coverage while significantly lowering the infrastructure required to serve it.

The model is available through Aikido’s Hugging Face organization and can be deployed on a node with four NVIDIA H200 GPUs using vLLM.

The company said it plans to optimize the model further, explore lower-bit formats, and fine-tune future versions for code analysis, remediation, tool use, and long-running security workflows.

Cut every SOC alert investigation by 21 min. Power your SOC with instant IOC context for immediate response: Integrate TI Lookup in your SOC