Aikido Launches Open-Weight Cybersecurity AI That Runs Locally

Aikido Security has introduced Altar, its first open-weight AI model for cybersecurity, with an emphasis on running defensive AI inside infrastructure controlled by the customer rather than sending sensitive code and security context to a third-party inference service.

The model is intended to power Aikido Machine, the company’s on-premises autonomous penetration-testing appliance. Aikido says the system can operate inside customer networks, including fully air-gapped environments.

What Aikido means by sovereign security intelligence

Aikido is framing Altar around a problem faced by organisations with strict data-residency or confidentiality requirements: advanced AI security tools can require source code, vulnerability context and infrastructure details to be processed by an external cloud model.

Running the model locally changes that trust boundary. Code and security findings can remain within infrastructure controlled by the organisation, which can be particularly relevant for regulated companies, critical infrastructure, government environments and teams working with proprietary software.

This does not automatically make a deployment secure. Local models still need access controls, patching, monitoring, model governance and careful handling of the systems they are allowed to test. The important difference is where inference happens and who controls the underlying environment.

Altar is open-weight, not simply a cloud API

Open-weight models make their trained model weights available under specified terms, allowing users to run the model on infrastructure they control. That is different from a hosted AI service where customers interact with a model only through an external API.

Aikido says Altar was compressed for practical deployment while retaining security reasoning needed for its defensive workflows. The company is positioning the model as a foundation for autonomous security testing rather than a general-purpose chatbot.

Organisations evaluating it should still examine the licence, hardware requirements, supported workflows and operational safeguards rather than assuming that “open-weight” means unrestricted use or zero deployment cost.

How it fits with Aikido Machine

Aikido Machine was introduced earlier as an on-premises AI pentesting system that can continuously test an organisation’s attack surface without sending its code outside the network. Altar provides the AI layer for that approach.

The intended workflow is defensive: identify weaknesses, validate whether they are exploitable and provide evidence that security teams can use for remediation. Running the model and testing appliance locally is designed for organisations that cannot send sensitive application context to external inference services.

Why local cybersecurity AI matters

Security tools often require unusually sensitive information. Source code can reveal intellectual property, credentials can expose systems if mishandled, and vulnerability data can become dangerous when leaked before remediation.

For some organisations, those concerns make external AI services difficult to approve even when the services have strong enterprise controls. A locally deployed model offers another architecture: the organisation controls the compute environment, network access and retention policies.

There are trade-offs. Running capable models locally can require specialised hardware and operational expertise. Organisations also become responsible for maintaining the inference stack and ensuring that the model is updated appropriately.

What security teams should evaluate

Teams considering locally deployed security AI should begin with their threat model rather than the novelty of the model. Important questions include what code and credentials the system can access, whether testing can affect production systems, how findings are stored and who can authorise autonomous actions.

They should also establish limits on what the AI can do. Automated security testing can be valuable, but unrestricted execution against sensitive systems can introduce operational risk. Human review, scoped permissions, audit logs and safe testing environments remain important.

Open weights do not remove security responsibilities

Keeping inference local reduces one category of data exposure, but it does not eliminate the need for conventional security controls. A compromised local server, excessive permissions or poorly isolated testing environment can still expose sensitive information.

Model outputs also need validation. AI systems can misunderstand context, produce false positives or suggest actions that are inappropriate for a particular environment. Security teams should treat autonomous findings as evidence to investigate, not as infallible conclusions.

A different deployment model for defensive AI

The model’s practical impact will depend on deployment requirements, accuracy and adoption. For now, the confirmed development is that Aikido has released its first open-weight security model and is using it to extend an on-premises, air-gapped approach to AI-assisted defensive testing.