Data Loss Prevention for LLM Traffic: Where It Has to Sit

Inspecting and controlling sensitive data before it leaves your perimeter in a prompt, tool call or model response. It differs from classic DLP because there is no file and no unsanctioned destination - the payload is JSON inside a TLS session to an API you pay for. Controls must understand the request schema and sit at an authenticated egress chokepoint.

Three reasons compound. The destination is allowlisted by design, so destination rules never fire. The content sits in a nested conversation array rather than a file or form field, which document-oriented engines parse poorly. And without TLS inspection there is nothing to parse - while even an inspecting proxy cannot tell which caller sent the request.

It depends on the detector. TrueFoundry’s built-in PII / PHI Detection is mutate-only - it always redacts, never rejects. For a hard block you need a validate-capable integration such as AWS Bedrock Guardrails, Azure PII or CrowdStrike. “PII detection” in a feature list never tells you whether the outcome is a redaction or a 400.

Yes. TrueFoundry runs in your VPC, on-prem, air-gapped, or hybrid, so prompts and responses never leave your domain even as you route across many providers.

Yes. It includes an MCP Gateway, an Agent Gateway, and an MCP & Agents Registry with tool-level access control. Agents on LangGraph, CrewAI, AutoGen, or a custom framework can all be governed centrally.

Oui. La passerelle est compatible OpenTelemetry et s'intègre à Grafana, Datadog, Prometheus, ou à votre pile technologique préférée. Elle trace chaque requête, du prompt à l'exécution de l'outil et du modèle, vous offrant ainsi une journalisation unifiée sans avoir à remplacer ce que vous utilisez déjà.