Should Your Company's Source Code Ever Leave Its Network for AI?

Belgian cybersecurity company Aikido has introduced Altar, an open-weight cybersecurity model designed to run inside customer-controlled infrastructure. It powers an on-premises security product intended for organisations that do not want source code, architecture documents or security findings sent to an external inference provider.

The announcement arrives as developers and security teams are asking harder questions about where AI tools process sensitive material. The ZCode repository-indexing incident shows how a useful coding feature can create a serious trust problem when its data movement and permissions are not clearly understood.

Cloud AI and local AI

With cloud AI, prompts and supporting context are sent to infrastructure operated by an external provider. This can be convenient and highly scalable, but organisations need contractual, technical and operational controls governing what leaves the network and how it is handled.

Local AI runs the model within infrastructure controlled by the customer. Local inference means the actual model computation happens there rather than through an external model API. This can help meet requirements around confidentiality, data residency or air-gapped environments.

What open-weight means

An open-weight model makes its trained parameters available under specified licence terms, allowing an organisation to deploy and operate the model itself. Open weights are not the same as fully open-source software, and they are not automatically safe. Teams still need to examine licensing, model provenance, updates, access controls and the surrounding application.

Aikido says Altar is a compressed and customised derivative of Z.ai's GLM-5.3 model. Its significance here is architectural: the model can be deployed locally as part of Aikido Machine rather than requiring customer code to be sent to a third-party inference service.

Enterprise AI architecture is a security decision

Where a model runs changes the trust boundary. Local deployment can reduce exposure to outside providers, but it does not remove the need for secure configuration, monitoring, patching and governance. Cloud deployments can also be appropriate when permissions, retention and isolation are well controlled.

The right question is not whether local or cloud is always better. It is which architecture matches the sensitivity of the data, the organisation's risk tolerance and the operational controls it can maintain.

Aiexl. takeaway

Where AI runs can be as important as what the AI can do. Source-code privacy, inference location and permission design should be explicit parts of an enterprise AI architecture—not details discovered after deployment.