Aikido Security has introduced Altar-1, a new open-weight AI model built to run entirely within an organization’s own infrastructure. This means source code, internal documents, and vulnerability reports stay inside the firewall — avoiding cloud inference and protecting data and compliance obligations. The announcement targets sensitive environments like banking, healthcare, and industrial systems that require airtight control over their data.
Altar-1 powers Aikido Machine, the company’s autonomous penetration-testing appliance. Previously, this system carried out continuous sweep and exploit operations across an organization’s attack surface. With Altar-1 added to the workflow, the model handles sensitive technical inputs locally, removing the need to rely on external AI services.
From GLM-5.3 to a slim, security-focused workhorse
The foundation of Altar-1 is a heavily modified version of the GLM-5.3 model, originally created by Z.AI. Aikido applied a multi-step reduction process: first using AWQ INT4 quantization to shrink the model, then employing expert pruning — trimming away specialists (experts) that contribute less to security assessment tasks. After pruning, Altar-1 stores 32.8% less data than the already quantized GLM-5.3 and drops from 1.51 TB to 328 GB overall.
Still, it retains critical capabilities. Of the 256 routed experts per backbone layer in the original model, 168 are preserved. The experts removed were chosen via benchmark-based analysis using internal pentest traces, not actual customer data. Aikido also added multilingual capability to ensure the model can review technical documentation, code, and interface text in different languages. A pruning method called Router-weighted Expert Activation Pruning (REAP) helped pick which expert modules to keep based on router activation patterns and significance during output computation.
Performance & deployment trade-offs
In internal tests spanning 32 known vulnerabilities in 30 repositories, Altar-1 achieved an average recall of 60.4%, rediscovering 23 of 32 issues over three runs. The non-pruned, quantized version of GLM-5.3 matched that same 23 vulnerabilities, reaching 61.5% recall. The full-precision model did slightly better (65.6%) and found 25 vulnerabilities. In effect, Altar-1 retains about 92% of the full model’s vulnerability coverage while cutting infrastructure needs significantly.
For deployment, Altar-1 is available via Aikido’s Hugging Face organization. Organizations can run it in-house on hardware with four NVIDIA H200 GPUs using vLLM. Going forward, Aikido plans to push further: exploring lower bit-precision formats and fine-tuning versions tailored to handle code analysis, remediation, tool orchestration, and prolonged security workflows.
Why this matters: AI-powered penetration testing has gained traction, but many organizations worry about security or compliance risks when sending internal information outside. Altar-1 addresses that directly by keeping everything inside the perimeter while delivering strong recall on vulnerability detection. It represents a move toward “sovereign” AI models that respect operational isolation.
