Data Loss Prevention

Overview

The Allternit LLM gateway includes built-in Data Loss Prevention (DLP) middleware that scans request and response bodies for secrets, credentials, and prompt-injection attempts. DLP runs after virtual-key authentication and rate limiting, but before the budget check and provider routing, so sensitive data is caught before any spend occurs.

How DLP works

The gateway buffers the request body (up to 16 MB).
Pattern matchers scan text for known secret types and injection heuristics.
Each match is assigned an action: block, redact, or warn.
Blocked requests return a content_policy_violation error and are recorded in usage events with status dlp_blocked.
Redacted requests continue, with matched secrets replaced by placeholders.
Warned requests continue, with a warning header returned to the caller.

Built-in secret patterns

Pattern	Example match
AWS access key	`AKIA...`
AWS secret key	Base64 secret when “aws” appears nearby
GitHub token	`ghp_...`, `gho_...`, `github_pat_...`
Provider API keys	Virtual keys and third-party provider API keys
Slack token	`xoxb-...`, `xoxp-...`
PEM private key	`-----BEGIN PRIVATE KEY-----`
JWT	`eyJ...` JSON Web Tokens
US Social Security number	`123-45-6789` (with SSA validation)
Credit card number	16-digit numbers passing Luhn validation

Secrets are never logged in plaintext. The gateway records only the pattern id and a SHA-256 hash of the matched bytes.

Injection heuristics

The DLP layer also scores messages for prompt-injection and jailbreak attempts:

Heuristic	Weight	Description
Instruction override	2	Phrases that attempt to override system instructions.
Chat markup delimiter	3	Delimiters such as `<	im_start	>`,` [INST]`, or similar sequence-breakers.
Base64 blob	2	Base64-encoded blobs longer than 400 characters that decode successfully. Capped at 4.

Default thresholds:

Action	Threshold
Warn	score ≥ 2
Block	score ≥ 5

Actions

Action	Behavior	HTTP response
`block`	Reject the request.	`400` with `allternit.content_policy_violation`
`redact`	Replace matched secrets with `[REDACTED]`.	Continues with `x-allternit-dlp: redacted`
`warn`	Allow the request and add a warning header.	Continues with `x-allternit-dlp: warned`

Configuration

Administrators can manage custom DLP rules through the LLM gateway admin API. The default action for built-in patterns is controlled by the environment variable ALLTERNIT_DLP_DEFAULT_ACTION (block, redact, or warn).

Environment variables

Variable	Values	Default	Description
`ALLTERNIT_DLP_DEFAULT_ACTION`	`block`, `redact`, `warn`	`block`	Default action for built-in pattern matches.
`ALLTERNIT_DLP_INJECTION_WARN_THRESHOLD`	integer	`2`	Score at which to warn on injection heuristics.
`ALLTERNIT_DLP_INJECTION_BLOCK_THRESHOLD`	integer	`5`	Score at which to block on injection heuristics.

Admin DLP rules

Organization admins can list and update custom rules at /api/v1/gateway/dlp/rules.

List rules

curl https://api.allternit.com/api/v1/gateway/dlp/rules \
  -H "Authorization: Bearer $ALLTERNIT_TOKEN"

{
  "rules": [
    {
      "id": "rule_abc123",
      "name": "custom-internal-token",
      "pattern": "internal_[a-z0-9]{32}",
      "action": "block",
      "enabled": true
    }
  ],
  "default_action": "block"
}

Create or update a rule

curl -X PUT https://api.allternit.com/api/v1/gateway/dlp/rules \
  -H "Authorization: Bearer $ALLTERNIT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "custom-internal-token",
    "pattern": "internal_[a-z0-9]{32}",
    "action": "block",
    "enabled": true
  }'

Response-side DLP

For non-streaming chat completions, the gateway can also scan the assistant response. When response-side redaction occurs, the header x-allternit-dlp-response: redacted is included.

Error codes

HTTP status	Code	Meaning
400	`allternit.content_policy_violation`	Request blocked by DLP or safety classifier.
400	`allternit.invalid_request`	Malformed DLP rule.
403	`allternit.insufficient_role`	Caller is not an organization owner or admin.
404	`allternit.not_found`	DLP rule not found.
500	`allternit.internal_error`	Internal gateway error.

Security overview — security model and controls
Inference hooks — custom pre/post-processing hooks
Admin gateway — LLM gateway administration endpoints

Get Started

Core Concepts

Gizzi Code CLI

Tools & ACI

Security & Enterprise

Admin API

BYOC Runtime

Surfaces

CI CD

Guides

Cookbooks

Release Notes

Overview

How DLP works

Built-in secret patterns

Injection heuristics

Actions

Configuration

Environment variables

Admin DLP rules

List rules

Create or update a rule

Response-side DLP

Error codes

​Overview

​How DLP works

​Built-in secret patterns

​Injection heuristics

​Actions

​Configuration

​Environment variables

​Admin DLP rules

​List rules

​Create or update a rule

​Response-side DLP

​Error codes

​Related pages

Overview

How DLP works

Built-in secret patterns

Injection heuristics

Actions

Configuration

Environment variables

Admin DLP rules

List rules

Create or update a rule

Response-side DLP

Error codes

Related pages