OPERATIONS & SAFETY
Connect external models. Check security.
AIDA checks the security when integrating OpenAI, Anthropic, and other model providers. The crucial factor is which data is processed for which task and via which access point.
Discuss your use case ↗
View larger image ↗
Examine the specific use
A provider name alone does not describe a data path. Business APIs, additional tools, and different contract models can have varying terms and conditions. Therefore, the review focuses on the actual intended connection and the selected functions.
Clarify data scope and confidentiality
Before any data transfer, it is determined which content is required and which information should remain local. Appropriate restrictions and permissions are agreed upon for sensitive content. Whether processing or anonymization is sufficient must be assessed for each specific use case.
Consider processing and storage
The review considers available information regarding processing location, storage, training usage, and involved services. These conditions may vary depending on the product, function, and contract. The current documentation for the specific access is authoritative.
Coordinate technical connection and permissions
Access to the model, permitted tools, and the handling of access credentials are part of the security framework. Logging, network paths, and potential write-back actions are also considered. The specific scope of the audit and the implemented controls are defined for the project.
Make the approval process traceable
The result of the review is a reasoned decision regarding the intended use: which tasks and data are authorized, which restrictions apply, and when a further review is necessary. Such authorization is not a blanket security certification of a provider.
Example: a technical summary
In an illustrative scenario, confidential drawings are stored locally. An external model is to be added for a general text draft. First, it is checked whether the task can be completed without the confidential drawing content. Only the agreed-upon data set is allocated for the external step.