AI enablement changes the organization's data boundary.

Classify data before sending it to an external model provider.

Only approved providers should receive company data.

AI should not autonomously make high-impact legal, financial, employment or security decisions.

Retrieved RAG documents are untrusted input and should not override system instructions.

Store secrets in environment variables or a secrets manager.

Never place credentials in:

• repositories
• prompts
• indexed knowledge
• frontend code

Useful metadata includes:

• workflow ID
• provider
• model
• prompt version
• timestamp
• review requirement

Avoid logging unnecessary sensitive prompt content.

Self-hosting does not automatically make AI secure.

It still requires:

• patched hosts
• authentication
• authorization
• TLS
• network segmentation
• monitoring
• access logging