The definition
A security layer between your staff and AI models. It inspects every prompt, detects sensitive content, applies policy and decides whether a request is allowed, masked, blocked or processed locally.
The difference from the layers that already exist, in three sentences: firewalls protect networks. DLP protects files. An AI Interaction Firewall protects what staff type into AI — and because it sits in front, it also decides which model is used for what.
What it does, in the order it does it
- Hard constraints. If the conversation is already sensitive, the mode is local only, or there is no cloud access, processing stays local. Image content is extracted locally and checked like text.
- Detection. Regex, a local NER model and a context-based scan run together, covering 23 PII categories across seven groups, up to the special categories under Art. 9 GDPR.
- Decision. No finding means cloud allowed. A finding with policy "block" forces local processing; with policy "anonymize" the request goes out masked.
What it is not
- Not an endpoint tool. It does not hang off a managed device or a browser extension.
- Not a gateway in someone else's network. It runs where you put it.
- Not a governance tool without enforcement. A policy that does nothing at runtime is a document.
Terms that mean the same field: AI TRiSM, AI Runtime Inspection and Enforcement, AI DLP, AI firewall. They are not identical, but they aim at the same gap.