Aikido launches cyber AI designed to inspect sensitive code without sending it to the cloud

Belgian cybersecurity company Aikido Security has launched an open-weight AI model designed to analyse proprietary software while keeping sensitive code inside an organisation’s own infrastructure.

The model, called Altar, is a customised and compressed version of Chinese AI company Z.ai’s GLM-5.3.

Unlike cybersecurity tools that require customers to send code to an externally hosted AI service, Altar can be deployed locally.

That means companies can use AI-assisted security analysis while retaining control over source code that may contain commercially sensitive intellectual property, security information or other confidential material.

Keeping proprietary code inside the organisation

The ability to run the model locally is central to Aikido’s proposition.

AI models are increasingly being used to identify vulnerabilities, review software and assist developers with security work.

But using externally hosted models can create an additional consideration for organisations that do not want proprietary code transmitted outside their own systems.

Altar is intended to provide an alternative by allowing the AI model to operate within infrastructure controlled by the customer.

Reuters reports that the technology will be incorporated into Aikido products used by customers including Belgian bank Belfius.

Built from GLM-5.3

Altar is based on Z.ai’s GLM-5.3, but Aikido has customised and compressed the underlying model for cybersecurity applications and local deployment.

The distinction between the two models is important when considering performance claims.

The US government’s Center for AI Standards and Innovation, or CAISI, recently described the underlying GLM-5.3 as the most cyber-capable open-weight model it had evaluated.

On CAISI’s aggregate measure, however, GLM-5.3 remained approximately four months behind the current US frontier in cyber capability.

Those findings concern GLM-5.3 itself and should not automatically be treated as equivalent benchmark results for Aikido’s modified Altar model.

Local AI expands cybersecurity options

Open-weight models allow companies to modify and deploy AI systems in ways that are more difficult with models available only through external services.

For cybersecurity teams, local deployment can be particularly useful when the material being analysed includes proprietary source code or other sensitive technical information.

Aikido’s launch therefore adds another deployment option as AI becomes increasingly integrated into software security.

Rather than requiring sensitive corporate code to travel to the AI, Altar is designed to bring the AI to the code.

Sources

• Reuters: Belgium’s Aikido launches cybersecurity AI model as demand for local tools grows
• NIST / CAISI: Assessment of Z.ai’s GLM-5.3 cyber capabilities