Aikido Security Launches On-Premises AI for Vulnerability Discovery
Aikido Security has released Altar-1, an open-weight AI model designed to run cybersecurity tasks like vulnerability discovery and penetration testing locally, preserving data privacy.
Aikido Security has introduced Altar-1, a novel open-weight artificial intelligence model engineered to execute defensive cybersecurity operations entirely within an organization's own infrastructure. This innovation aims to empower security teams to leverage advanced AI for tasks such as vulnerability discovery and penetration testing without the necessity of transmitting sensitive source code, internal documentation, or security findings to external cloud-based inference services. Altar-1 is specifically tailored for organizations grappling with stringent privacy, regulatory, or operational mandates, including financial institutions with data-residency obligations, healthcare providers handling protected patient information, and industrial entities operating isolated or air-gapped operational technology (OT) environments. By ensuring both the AI model and the associated data remain under the direct control of the customer, Aikido is positioning Altar-1 as a sovereign security intelligence solution.
The Altar-1 model serves as the core intelligence engine for Aikido Machine, the company's autonomous pentesting appliance. Aikido Machine is architected to operate seamlessly within customer environments, continuously identifying, exploiting, and validating weaknesses across an organization's attack surface. The integration of Altar-1's local AI reasoning capabilities enhances this workflow, enabling security testing to process highly sensitive technical context without the risk of transmitting it to third-party entities.
Aikido developed Altar-1 by significantly compressing and optimizing Z.AI's open-weight GLM-5.3 model. The model's storage footprint was reduced from an initial 1.51 terabytes to a more manageable 328 gigabytes. This was achieved through a two-stage process: first, applying AWQ INT4 quantization, which reduced the model size to 488.2 GB, followed by expert pruning to remove components deemed less critical for targeted security workloads. Mixture-of-experts (MoE) models, like GLM-5.3, contain numerous specialized neural network components, or 'experts.' While efficient during inference by selecting only a subset of experts per token, the larger pool typically demands substantial memory.
Aikido reported preserving 168 of the original 256 routed experts within each backbone layer, effectively removing 88 experts. This pruning process resulted in a substantial 78.2 percent reduction from the original full-precision model and a 32.8 percent reduction from the already quantized checkpoint. The selection of experts to preserve was guided by traces derived from internal penetration testing benchmarks, rather than customer data. These traces encompassed representative code snippets, tool calls, and agent responses generated during simulated security testing workflows.
To ensure comprehensive language understanding capabilities, Aikido incorporated multilingual text during the model's calibration phase. This allows Altar-1 to effectively review documentation, business rules, application behaviors, and user interfaces across different languages. The company utilized Cerebras REAP (Router-weighted Expert Activation Pruning) to intelligently select experts based on router weights and output magnitude, thereby preserving critical reasoning capabilities in coding, cybersecurity, and natural language processing.
In internal benchmarking tests involving 32 known vulnerabilities across 30 repositories, Altar-1 demonstrated a strong performance, achieving an average recall of 60.4 percent per run and successfully rediscovering 23 of the 32 vulnerabilities at least once across three separate runs. The quantized but unpruned GLM-5.3 model achieved a slightly higher recall of 61.5 percent and found the same 23 vulnerabilities. For comparison, the full-precision parent model achieved a recall of 65.6 percent and identified 25 vulnerabilities.
Aikido asserts that Altar-1 successfully retains approximately 92 percent of the full model's vulnerability coverage while drastically reducing the infrastructure requirements for deployment. The model is accessible via Aikido's Hugging Face organization and can be deployed on a server equipped with four NVIDIA H200 GPUs, utilizing vLLM for inference. The company has indicated plans for further optimization, including exploring lower-bit formats and fine-tuning future versions for enhanced code analysis, remediation, tool utilization, and extended security workflow automation.
This release addresses a critical need for organizations that cannot or will not send sensitive data to cloud-based AI services due to regulatory compliance, data sovereignty concerns, or operational security requirements. By offering a powerful, locally deployable AI model for offensive security tasks, Aikido Security is enabling a new era of private, sovereign cybersecurity intelligence.