Aikido Security has introduced Altar-1, its first open-weight security model, purpose-built for deployment within customer-controlled and air-gapped environments. This innovative solution, a highly compressed version of Z.AI's GLM-5.3, provides powerful AI capabilities for autonomous pentesting. Altar-1 is integral to Aikido Machine, the company's appliance for on-premises and air-gapped network security, with its public weights available on Hugging Face, deployable via vLLM on a single node equipped with four NVIDIA H200 GPUs.
Addressing Secure AI Deployment Challenges
Altar-1 directly addresses critical challenges in deploying AI for security:
-
Data Residency: Traditional closed frontier models transmit sensitive security context—source code, architectural data, findings—outside an organization's secure network. This is untenable for banks with strict data mandates and OT environments lacking internet.
-
Open-Weight Deployment Hurdles: While open-weight models solve data residency, their substantial size creates obstacles. Large Mixture-of-Experts (MoE) models store all experts, and extensive security agent context competes with model weights for limited GPU memory.
Altar-1's Compact Engineering: Quantization and Pruning
Altar-1's efficiency stems from a sophisticated two-step compression applied to GLM-5.3, an MoE model originally featuring 753 billion parameters with 40 billion active per token. The process began with quantization using the cyankiwi GLM-5.3-AWQ-INT4 checkpoint, storing expert weights in 4 bits (W4A16) while critical components maintained BF16 precision.
Subsequently, Aikido employed Cerebras REAP (Router-weighted Expert Activation Pruning) to optimize the model further. REAP scores experts by router weight and output magnitude, enabling the removal of 88 experts (34.4%) from the original 256 per layer, retaining 168 without retraining. Calibration used Aikido’s pentesting harness traces and public datasets, avoiding customer data. The final Altar-1 model boasts a storage footprint of 328.0 GB, a 78.2% reduction from the original 1,506.7 GB BF16 GLM-5.3, and 32.8% smaller than its 488.2 GB AWQ quantized parent. Fidelity tests indicated a KL divergence of 0.506 nats against the full BF16 model.
Performance, Deployment, Licensing, and Future
Internal CVE benchmark tests across 32 known vulnerabilities showed Altar-1 maintained strong performance. Pruning resulted in approximately one-point recall reduction and no loss in vulnerability coverage compared to its AWQ checkpoint. Against the original parent model, Altar-1 retained 92% of covered vulnerabilities, with a 5.2-point lower recall. While the benchmark's scope is targeted, a vendor-reported instance highlighted Altar-1's success in identifying a critical-severity vulnerability during a client's production pentest.
For deployment, Altar-1 requires Hopper GPUs (H100 or H200). The 328 GB model weights across four H200 GPUs provide ample space for a 128k-context KV cache at production batch sizes. The model operates under the GLM-5.3 License, permitting commercial use, modification, and redistribution; however, Model-as-a-Service operators generating over $10 billion in annual revenue must undergo a Z.AI security review. Altar-1 is an open-weight model, not an OSI-approved open-source solution. Aikido plans to explore lower-bit formats and develop fine-tuned models for specialized security workflows, expanding Altar-1's capabilities further.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
