Pete Finnigan's Oracle Security Weblog

One area I have not covered in previous blogs about Oracle Security and AI is sovereignty of that AI and of course in particular of the data. This means that the control of the hardware (compute and GPU) and the software (the model software and weights) and of course the data that is passed to the models and used by the models is controlled and restricted. Providers try and control all of these aspects so that the data and compute are within the borders of the country.

Oracle offers sovereign AI in terms of its OCI cloud offerings in terms of public cloud and even using OCI cloud in the customers data center.

Why do we need sovereign AI? because if you just used public cloud AI then your data could be stored and used by the AI and in effect you are losing your data to the cloud based AI. So the idea of sovereign AI is to protect your data and company secrets.

You can use commercial sovereign solutions such as Oracle OCI but depending on your needs using a local LLM will offer the best solution as the LLM and hardware and all prompts and data passed to it remain in your server room and on your own network. This is the much better approach as no data leaves your premise. Yes, you do not get the latest Claude or Gemini or Chatgpt BUT a local LLM with RAG and all your documents and web search enabled is a good way to go.

In general there is no audit trail in the same sense as the Oracle database in AI or with AI usage.

But it is obvious; do not paste company data or secrets into a chat interface as this will end up in the cloud. So you need to know if any employee tries this.

Oracle security is all about knowing where the data is and who can access that data and how is that data protected. We also must combine audit trails in the Oracle database with the security and protection around the data that matters. We must not only audit access to the data but also changes to security such as adding or changing users, roles, profiles, privileges

In AI, if we use AI then we must also know what data is passed to AI and how; we also must know if the settings of the AI have changed and when and by who.

Using a local LLM with suitable RAG to fill the gaps between the frontier models and the open free models available for local hosting will allow you to host the AI locally and therefore prevent your data from leaving the premises and not ending up in a cloud based AI. If we cannot do this then use a sovereign AI solution.

Whether the solution is commercial or free most AI solutions have some level of audit controls so that changes to settings, logon/logoff, rag, prompts and results can be audited. For instance AnythingLLM out of the box audits (stores) conversations, workspace history, documents added, users, chats but does not provide an immutable audit trail for configuration changes, plugins or model changes. BUT, because it is open source this could be added.

Bottom line; consider sovereign AI to ensure that your data / code / PII / GDPR / Business logic is not leaked or exfiltrated. Also look at auditing all LLM use in your organisation (not cost of tokens) so that you know what changed and when and what was sent to the LLM and what was returned.

And remember that if you capture prompts in the audit trail then these potentially contain data that should be protected; so protect the AI audit trail.

#oracleace #sym_42 #ai #llm #sovereign #aisecurity