> ## Documentation Index
> Fetch the complete documentation index at: https://docs.rubric-app.com/llms.txt
> Use this file to discover all available pages before exploring further.

# DLP

> Detector configuration and the detection result shape.

The SDK ships with a regex-based DLP detector and an optional Presidio-backed one (via the `[presidio]` extra). DLP runs **before** policy evaluation so detections become first-class policy fields.

See [DLP concepts](/concepts/dlp) for the philosophy and policy patterns. This page is the API.

## Enable

Three knobs — pick one:

```python theme={null}
from rubric import Governance

# 1. Constructor kwarg
with Governance.bootstrap(agent_name="…", dlp="auto") as gov:
    ...

# 2. Constructor with a concrete detector
from rubric.dlp import RegexDetector
with Governance.bootstrap(agent_name="…", dlp=RegexDetector()) as gov:
    ...

# 3. Env var
# AG_DLP=auto python my_agent.py
```

## Modes

```python theme={null}
from rubric.constants import (
    DLP_MODE_OFF,
    DLP_MODE_REGEX,
    DLP_MODE_PRESIDIO,
    DLP_MODE_AUTO,
)
```

| Constant            | Value        | Behavior                                            |
| ------------------- | ------------ | --------------------------------------------------- |
| `DLP_MODE_OFF`      | `"off"`      | Detector disabled.                                  |
| `DLP_MODE_REGEX`    | `"regex"`    | Built-in regex patterns only.                       |
| `DLP_MODE_PRESIDIO` | `"presidio"` | Presidio NER + regex (requires `[presidio]` extra). |
| `DLP_MODE_AUTO`     | `"auto"`     | Presidio if available, otherwise regex.             |

## DlpDetection

```python theme={null}
class DlpDetection(BaseModel):
    detected: bool
    severity: Literal["low", "medium", "high"]
    types: list[str]
    matches: list[DlpMatch]
```

| Field      | Notes                                                                            |
| ---------- | -------------------------------------------------------------------------------- |
| `detected` | `True` if any pattern fired. `False` only if the detector ran and found nothing. |
| `severity` | The highest severity tier of any match.                                          |
| `types`    | Sorted list of matched type names (`EMAIL`, `SSN`, `AWS_KEY`, …).                |
| `matches`  | Detailed match records — value (sometimes redacted), location, type, severity.   |

## DlpMatch

```python theme={null}
class DlpMatch(BaseModel):
    type: str
    value: str          # redacted to first/last 2 chars when severity is high
    severity: Literal["low", "medium", "high"]
    location: str       # JSON path in the scanned payload
```

## Custom detector

Any object with this shape:

```python theme={null}
class MyDetector:
    def detect(self, payload: dict[str, Any]) -> DlpDetection | None:
        ...  # return None for no detection, DlpDetection for a hit
```

Pass via `dlp=MyDetector()` on bootstrap. The SDK calls `.detect()` once per `evaluate()`, passing the combined `metadata` dict (input + args + kwargs).

## Built-in regex patterns

The default regex detector covers:

| Type                    | Severity |
| ----------------------- | -------- |
| `EMAIL`                 | low      |
| `PHONE`                 | low      |
| `IP_ADDRESS`            | low      |
| `JWT`                   | medium   |
| `GITHUB_TOKEN`          | medium   |
| `SLACK_TOKEN`           | medium   |
| `GENERIC_API_KEY`       | medium   |
| `SSN`                   | high     |
| `CREDIT_CARD`           | high     |
| `AWS_ACCESS_KEY`        | high     |
| `MEDICAL_RECORD_NUMBER` | high     |

Override per-detector by subclassing and changing the `_PATTERNS` map.

## Failures

If a detector raises during `detect()`, the SDK logs and treats it as no-detection. Your `evaluate()` proceeds with `dlp_detected: false`. Fix the detector and you're back. Detector exceptions never propagate up to the caller.