A data-loss-prevention server can now veto a prompt before the model sees it
Anthropic shipped inline DLP with an open webhook that points at existing platforms. Putting the enforcement point in front of the model, rather than behind it, is the correct architecture and it took two years.
Anthropic has shipped inline data loss prevention: a DLP server can veto a prompt before Claude sees it, via an open webhook protocol that can point at a company's existing DLP platform.
The ordering is the whole story. Nearly every enterprise AI control built so far inspects the output — scan the response, redact what leaked, log it. That is auditing after the fact, and by then the sensitive material has already been transmitted to a third party.
An inline veto moves the enforcement point in front of the model. The prompt does not leave until the policy engine says it may. That is how every other egress control in a regulated environment already works, and it is the shape security teams have been asking for since the first enterprise deployment.
Pointing at existing DLP platforms rather than shipping a proprietary classifier is the other correct decision. Companies have spent years tuning those rules against their own document estate; a vendor-supplied model deciding what counts as sensitive would have to relearn all of it and would be wrong in ways nobody could audit.
The cost is latency and a new failure mode: if the DLP server is down, does the prompt fail open or closed? Fail-open defeats the control, fail-closed makes an external dependency into an availability risk for the assistant. That choice belongs to the customer and should be configurable.
It lands in the same week as a protocol for moving user context between vendors. Both are about where a boundary sits and who controls it.
TechManiacs — AI Security Daily Briefing: August 14, 2026 → · The Sysadmin Desk — Enterprise AI Weekly: August 3–8, 2026 →