Aikido releases open-weight AI cybersecurity model, Altar

Belgian cybersecurity startup Aikido has launched an AI model, Altar, specifically designed for security operations.

The release comes as some organisations seek AI solutions that can be deployed fully within their own infrastructure, responding to concerns about sharing sensitive code and operational data with external providers.

Altar is an open-weight model developed to support defensive cybersecurity tasks without requiring organisations to send internal context or source code outside their own environments.

According to Aikido, the model can be installed and operated entirely within customer-controlled systems, including entirely air-gapped networks.

The company describes Altar as a next step building on its existing autonomous penetration testing tool, Aikido Machine, which also runs on local infrastructure.

By equipping this appliance with Altar’s capabilities, Aikido aims to offer services such as code vulnerability detection and validation while maintaining full data residency.

Aikido Altar is based on Z.AI’s GLM-5.3 model but has been substantially reduced in size to address the high resource demands of advanced AI systems.

The original model occupied 1.51TB, but through quantisation and expert pruning, this was reduced to 328GB.

The company states this represents a 78.2% size reduction compared with the full-precision model and a 32.8% reduction relative to an already quantised checkpoint.

In internal evaluation, Aikido tested Altar on 32 known vulnerabilities across 30 repositories, running three assessments per case.

According to these results, Altar achieved 60.4% recall per run and re-identified 23 of the 32 vulnerabilities at least once, retaining 92% of the parental model’s coverage while cutting storage requirements by a third.

The company notes that while recall per run dropped slightly compared with parent models, most vulnerability coverage was preserved.

The model was rolled out to Aikido’s customer systems immediately following evaluation and was credited with the identification of a critical-severity vulnerability during a client test.

Altar is available for deployment using a 4-H200s node and the latest version of vLLM.

The company said further developments are planned, including exploration of additional compression techniques and training tailored to security-specific workflows.