Back to News
ProductAI Understanding briefing

Aikido Security releases Altar-1, a pruned security-focused AI model

Aikido Security has launched Altar-1, an open-weight security model derived from Z.AI’s GLM-5.3, designed for on-premises autonomous pentesting.

4 min readRead the linked source
Source-provided image accompanying Aikido Security releases Altar-1, a pruned security-focused AI model
Source referenceSource recorded
Publisher
marktechpost.com
Source link
marktechpost.comhttps://www.marktechpost.com/2026/09/25/aikido-security-releases-altar-1-an-open-weight-security-model-pruned-from-glm-5-3-to-328-gb/
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Start here

Key terms

Mixture of Experts (MoE)
An architecture with specialized subnetworks where only selected experts run per input.
Memory (Agent Memory)
Stored context an AI agent uses across steps or sessions to improve continuity.
Structured Output
Model output constrained to a defined schema such as JSON, tool arguments, or typed fields.
Test yourselfAI Models Explained Quiz

What happened

Aikido Security has released Altar-1, an open-weight security model pruned from the 753B parameter GLM-5.3 model. The model is optimized for deployment on private, air-gapped, or on-premises infrastructure, specifically to support autonomous pentesting tasks. According to the report, the model has been reduced to 328 GB, allowing it to fit within the memory constraints of a 4x NVIDIA H200 GPU node while maintaining space for a 128k-context KV cache.

Aikido Security developed Altar-1 by applying two compression steps to the 753B parameter GLM-5.3 mixture-of-experts (MoE) model. The pruning process reduced the number of experts from 256 to 168 while maintaining the routing logic that selects 8 experts per token.

The model is designed to power the company's 'Aikido Machine' autonomous pentesting appliance. It is currently available on Hugging Face and is compatible with vLLM, utilizing the Marlin MoE kernel for inference.

Aikido reports that Altar-1 retains 92% of the covered vulnerabilities from its internal 32-CVE benchmark compared to the parent model, with a 5.2-point decrease in recall. The vendor also claims the model successfully identified a critical-severity vulnerability during a client's production pentest.

Source details: marktechpost.com ↗

Why it matters

Altar-1 addresses a critical deployment gap for organizations with strict data-residency requirements, such as banks and operational technology (OT) operators, who cannot send sensitive source code or architecture documentation to cloud-based frontier models. By providing a model that can run locally, Aikido enables these entities to perform autonomous security analysis without exposing proprietary data to external networks. The model's pruning process specifically preserves expert parameters related to coding and , which are essential for security-focused workflows. However, the model's performance is limited to targeted CVE rediscovery within specific pipelines, and its efficacy in blind discovery or exploit validation remains unproven outside of the vendor's internal benchmarks.

The primary utility of Altar-1 is enabling security analysis in environments where data residency is a legal or operational mandate. By moving the model on-premises, organizations avoid the risks associated with transmitting sensitive internal architecture and source code to third-party cloud providers.

The model's architecture is specifically tuned to balance memory usage with the long-running context requirements of security agents. By pruning the MoE structure, Aikido aims to solve the 'deployment gap' where memory-intensive models often struggle to maintain sufficient KV cache space alongside model weights.

The reliance on vendor-provided benchmarks means that the model's real-world performance in diverse, non-standardized security environments is not yet independently verified. The narrow scope of the current benchmark—focused on CVE rediscovery—suggests that the model is currently a specialized tool rather than a general-purpose security researcher.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Interactive Concept Check+10 Points
AI Models Explained Quiz

In AI, what are a model's "parameters"?

What to watch next

Users should note that the model requires high-end hardware, specifically Hopper-architecture GPUs (H100 or H200). While the model fits on 4x H200 GPUs, a 4x H100 80GB node (320GB total) is insufficient to hold the 328GB of weights, potentially limiting accessibility for some enterprise environments. Additionally, while the model is open-weight, it is not OSI-approved open source and remains subject to the original GLM-5.3 license, which includes specific restrictions for large-scale commercial operators. Future developments include potential fine-tuning for security workflows and experimentation with lower-bit formats like EXL3 to increase the number of active experts.

Hardware requirements are a significant barrier to entry. The 328GB weight footprint necessitates a 4x H200 configuration, which may be cost-prohibitive or physically unavailable for many potential users.

The licensing terms for Altar-1 include a clause requiring a security review from Z.AI for any Model-as-a-Service operator generating over $10B in annual revenue, which may impact large-scale commercial adoption.

Aikido has signaled plans to explore lower-bit formats like EXL3. This is a critical area to watch, as it could allow for higher expert retention and potentially improved reasoning capabilities in future iterations of the model.

Related guides & quizzes

AI Models ExplainedAI AgentsAI EthicsTest what you know — try a free AI quizLook up an AI term in our glossaryFollow the AI model release tracker
Found this useful?