Image: cdn.prod.website-files.com · rights & removal

Introducing Aikido Altar: the model that makes sovereign security intelligence possible

Reporting by Aikido Security ResearchRead the original at aikido.dev

Executive Summary

The introduction of Altar, an open-weight security model based on the autonomous pentesting appliance Aikido Machine, focuses on bringing frontier-grade defensive security into controlled infrastructure. This is achieved by running autonomous testing inside customer environments, including air-gapped settings, without transmitting sensitive context externally. Altar incorporates advanced AI to perform defense without sending organizational context to third parties. The development addressed a deployment gap where large open-weight models lack the reasoning quality of frontier models while remaining deployable. To bridge this, the process involved optimizing a powerful model like GLM-5.3 through quantization and expert pruning, reducing its size from 1.51 TB to 328 GB while retaining substantial reasoning capabilities. The methodology focused on workload-specific optimization using traces from pentesting harnesses to guide the removal of experts, balancing security performance with efficiency gains for agentic security workloads.

Facts Only

* Altar is the first open-weight security model.
* Aikido Machine is an autonomous pentesting appliance running inside customer infrastructure.
* The process involved using GLM-5.3 as a starting point, reducing its size to 328 GB via quantization and expert pruning.
* Model size reduction achieved 78.2% compared to the full precision model.
* A further 32.8% reduction was achieved by applying AWQ INT4 quantization to the already-quantized checkpoint.
* Expert pruning involved removing expert weight blocks based on traces from pentesting harnesses and multilingual text to preserve language understanding.
* The final Altar model retained 168 of the original 256 routed experts.
* An evaluation against an internal CVE benchmark showed Altar averaged 60.4% recall across three runs, covering 23 of 32 vulnerabilities.
* Quantized GLM-5.3 AWQ achieved 61.5% recall and covered the same 23 vulnerabilities.
* The original full-precision GLM-5.3 averaged 65.6% recall and covered 25 of 32 vulnerabilities.
* Altar retained 92% of the parent's vulnerability coverage at a 33% reduction in storage.

Full Take

The narrative presents a sophisticated argument for shifting frontier model capabilities to deployable, sovereign environments by framing model size and efficiency as direct constraints on security utility. The process highlights a tension between raw capability (frontier models) and operational necessity (data residency mandates). The technique of using workload traces from internal benchmarks to guide pruning expert selection suggests an attempt to create context-aware, domain-specific knowledge distillation, moving beyond simple statistical reduction toward functional relevance. However, the resulting performance metrics—where reduced size correlated with a slight decrease in raw recall compared to the full model—introduce the core tension: optimizing for operational constraints versus maximizing absolute accuracy. The ultimate implication is that sovereign security intelligence requires an architecture where efficiency is not a secondary optimization but a foundational prerequisite; if capability must be retained, the cost of deployment and inference must be reframed as an acceptable trade-off within controlled boundaries.
Bridge Questions: If the goal is to prioritize real-world defense workflows over benchmark recall, what new metrics should guide expert selection in future model compression? How can the system evolve from static pruning based on benchmarks to a dynamic, self-learning pipeline that adapts to evolving operational constraints? What are the long-term implications for defining the necessary "reasoning quality" required for high-stakes security decisions within restricted environments?

From the original · Aikido Security Research

Today, we’re introducing Altar, our first open-weight security model, built to bring frontier-grade defensive security into infrastructure you control. We took our first step toward sovereign security intelligence with our autonomous pentesting appliance Aikido Machine, which runs entirely inside a customer’s own infrastructure, including fully air-gapped environments.

Read the full story at aikido.dev

Sentinel — Human

Confidence

The text reads like a detailed technical white paper or an engineering blog post, exhibiting high specificity and coherent argumentation typical of expert communication rather than generic content generation.

Signals Detected

Sentence length variance is erratic; shifts between technical detail and philosophical framing.

Strong, focused argument structure linking technical constraints (memory/architecture) directly to a high-level goal (sovereign security).

The flow transitions smoothly from introduction (Altar) to problem framing (deployment gap), technical mechanics (MoE, pruning, quantization), and final implications.

Specific benchmark results (60.4% recall, 92% coverage) tied to internal testing provide grounding; the discussion of trade-offs is nuanced and avoids sweeping generalizations.

Human Indicators

The integration of highly specific technical details (e.g., W4A16, REAP contribution score) with abstract philosophical goals suggests deep domain expertise rather than pure synthesis.

The subtle acknowledgment of trade-offs—where pruning might lose language understanding versus code performance—indicates the complex negotiation inherent in real research.