AI Data ProtectionRouted traffic is checked before it reaches the model.
Applications you configure to use RedactWall send prompts through the gateway, and answers return through the same path. RedactWall checks both directions and stops what policy says should not move.
Typed into an AI assistant
Set up the direct deposit. Routing number: 555000005, checking account number 000123456789. Draft the confirmation letter.
- Read complete
- Decide complete
- Record complete
What the AI provider receives
Set up the direct deposit. Routing number: [REDACTWALL:ROUTING_NUMBER:1], checking account number [REDACTWALL:BANK_ACCOUNT:1]. Draft the confirmation letter.
Two hard stops in one message. The rest of the prompt went through.
The answer comes back the same way and is checked before staff see it.
Recorded on your evidence stream
A sample prompt with made-up values.
What that looks like in practice
An application routed through RedactWall sends customer data to an AI assistant
Stopped before it reaches the model, and the attempt is on the record.
The assistant repeats a member ID back in its answer
Caught on the way back, before your staff see it.
A developer points an internal application at a new AI provider
The same rules apply, with no code to rewrite.
The evidence record is briefly unavailable
RedactWall refuses the request rather than letting it through with nothing on the record.
What you get
One route per compatible application
Point a compatible application at RedactWall instead of the provider. The application must use that route for the gateway to govern its traffic.
The answer is checked too
Most controls watch what staff send. This one also checks what the model says back, which is where account numbers reappear.
Every decision is evidence
Allowed, redacted, blocked or held: each one lands on a record your own team can check.
Where this fits with the other surfaces
A staff member submitting through a supported AI site on an enrolled browser is covered by Device & Browser Protection, and a tool call routed through a RedactWall guard is covered by AI Agent Security.
Start with one application
For the proposed pilot, configure one compatible application to use the Anthropic gateway, then verify request and response decisions on that specific route.
- Only traffic routed through RedactWall is checked
- Prompts and responses are checked before forwarding
- Evidence records the decision, not inspected content