Data protection / Request + Response
DLP and LLM Guard run before the model call and before the response is returned. Policy determines whether to observe, redact, block or change the route.
Deterministic rules produce repeatable results and explain which detector matched. An organization can extend the set without changing code.
The request is inspected before the provider. The response is inspected before it reaches the user. Block stops traffic; redact replaces the match with the configured marker.
Kazakhstan identifiers and bank details as built-in detectors.
OpenAI API keys, AWS access keys, email addresses and common secret patterns.
Tenant dictionaries and constrained regex with severity, replacement and active state.
Guard does not claim perfect jailbreak detection. It returns a deterministic score and signal identifiers for threshold tuning.
Instruction override, prompt exfiltration, role impersonation, jailbreak markers, tool coercion, data exfiltration and encoded payloads.
Policy can move a request from a public alias to a pre-approved Private LLM. Activation will not allow a public provider as the sensitive target.
The same detector can produce different actions in test, general and restricted policies.
These limitations are important for designing a sound pilot.
| Scope | Status | Comment |
|---|---|---|
| Text | Available | Requests and responses in the OpenAI-compatible chat contract. |
| Files, OCR and images | Planned | Require a separate extraction and inspection pipeline. |
| Local NER | Extension | An extension interface is not the same as a delivered and evaluated model. |
| Guaranteed detection of every jailbreak | Not promised | Quality is measured on an agreed synthetic/red-team dataset. |
A pilot starts by agreeing data categories, an acceptable false-positive rate and observe mode.
Discuss a DLP pilot →