OpenAI Data Loss Prevention Policy

Overview | Try It | Make Reference

Overview

This policy was created with the Flex Gateway Policy Development Kit (PDK). To find the complete PDK documentation, see PDK Overview on the MuleSoft documentation site.

The policy has the following properties

Property	Description	Type
`presidio-analysis-service`	The Presidio Analyze service location running in Docker	String
`langauge`	The language used by Presidio in ISO-639-1 format	String
`score_threshold`	The score threshold in Presidio for it to be flagged as sensitive (0-1)	String
`entities`	An array of entities to look for in the OpenAI request	Array
`action`	Log - Log sensitive data but continue or Reject - if sensitive data found return 401 (Unauthorized)	String

Example

When calling the OpenAI API a user will potentially include sensitive data in the prompt.

However, with Flex gateway being used and openai.api.com as the upstream api we can intercept the request and use pii checking utilities to look for sensitive data. In this case Microsoft's Presidio.

This policy assumes you have already added your OpenAI API key to the request header. You can do this in your http client or use the policy Open API Key Management Policy

>  curl http://localhost:8081/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-3.5-turbo",
    "messages": [
      {
        "role": "system",
        "content": "You are an assistant, skilled in explaining MuleSoft concepts with creative flair. Keep responses free from bias and without obsenities"
      },
      {
        "role": "user",
        "content": "create a poem in less than 10 words about Max Mule"
      }
    ]
  }'

With the policy applied the API returns Unauthorized 401

Your OpenAI request has sensitive data:
PERSON at 42,50: with certainty 0.85

Try It

These setup steps use Flex Gateway in connected mode running locally on a Mac. Make sure you have docker desktop installed.

Flex Gateway

Step 1. Follow the instructions here to get Flex Gateway up and running with openai.api.com/v1 as the upstream endpoint.

Step 2. Install Microsoft Presidio using Docker

docker pull mcr.microsoft.com/presidio-analyzer

docker run -d -p 5001:3000 mcr.microsoft.com/presidio-analyzer:latest

This will run Presidio's analyzer using the default port 5001.

Step 3. Verify Presidio is running correctly by running the following curl command

curl -X POST http://localhost:5001/analyze -H "Content-type: application/json" --data "{ \"text\": \"John Smith drivers license is A123456\", \"language\" : \"en\"}"

If the service is running you will get the response that responds with a match for name and drivers licence.

[{"analysis_explanation": null, "end": 10, "entity_type": "PERSON", "recognition_metadata": {"recognizer_identifier": "SpacyRecognizer_140733976679664", "recognizer_name": "SpacyRecognizer"}, "score": 0.85, "start": 0}, {"analysis_explanation": null, "end": 37, "entity_type": "US_DRIVER_LICENSE", "recognition_metadata": {"recognizer_identifier": "UsLicenseRecognizer_140733983661648", "recognizer_name": "UsLicenseRecognizer"}, "score": 0.6499999999999999, "start": 30}]%

Step 3. Download this policy and use make build then make release.

Step 4. Apply the policy in API Manager to the API created in Step 1. Leave all default options set.

Step 5. With the policy applied with default settings any prompt data with name, drivers licence, credit card, email addresses or phone numbers will be rejected. Other options you can add include

CRYPTO
DATE_TIME
IBAN_CODE
IP_ADDRESS
LOCATION
MEDICAL_LICENSE
URL
US_BANK_NUMBER
US_ITIN
UK_NHS and more all documented here.

To test it you can use the following:

The following request should be okay since it contains no PII.

curl -X POST http://localhost:8081/flex-api/chat/completions   -H "Content-Type: application/json"  -d '{
    "model": "gpt-3.5-turbo",
    "messages": [
      {
        "role": "system",
        "content": "You are an assistant, skilled in explaining MuleSoft concepts with creative flair. Keep responses free from bias and without obsenities."
      },
      {
        "role": "user",
        "content": "Compose a poem in less than 10 words."
      }
    ]
  }'

The following request should fail since we added a name

curl -X POST http://localhost:8081/flex-api/chat/completions   -H "Content-Type: application/json"  -d '{
    "model": "gpt-3.5-turbo",
    "messages": [
      {
        "role": "system",
        "content": "You are an assistant, skilled in explaining MuleSoft concepts with creative flair. Keep responses free from bias and without obsenities."
      },
      {
        "role": "user",
        "content": "Compose a poem in less than 10 words about Max Mule."
      }
    ]
  }'

This should reply with the Unauthorized 401 response

Your OpenAI request has sensitive data:
PERSON at 43,51: with certainty 0.85%

Make command reference

This project has a Makefile that includes different goals that assist the developer during the policy development lifecycle.

For more information about the Makefile, see Makefile.

Setup

The make setup goal installs the Policy Development Kit internal dependencies for the rest of the Makefile goals. Since these dependencies are provided by the Anypoint Platform, it requires the user to be authenticated with a set of valid Anypoint credentials.

For more information about make setup, see Setup the PDK Build environment.

Build asset files

The make build-asset-files goal generates all the policy asset files required to build, execute, and publish the policy. This command also updates the config.rs source code file with the latest configurations defined in the policy definition.

For more information about creating a policy definition, see Defining a Policy Schema Definition.

For more information about make build-asset-files, see Compiling Custom Policies.

Build

The make build goal compiles the WebAssembly binary of the policy. Since the source code must be in sync with the policy definition configurations, this goal runs the build-asset-files before compiling.

For more information about make build, see Compiling Custom Policies.

Run

The make run goal provides a simple way to execute the current build of the policy in a Docker containerized environment. In order to run this goal, the playground/config directory must contain a set of files required for executing the policy in a Flex Gateway instance:

A registration.yaml file generated by performing a Flex Gateway registration in Local Mode. If you already have an instance registered in Local mode, you can reuse the registration file you have and copy it in the playground/config folder. Otherwise, to complete the registration we recommend using the Anypoint Platform:
1. Go to Runtime Manager
2. Navigate to the Flex Gateway tab
3. Click the Add Gateway button
4. Select Docker as your OS and copy the registration command replacing --connected=true to --connected=false.
5. Paste the command and run it in the playground/config directory.
An api.yaml file updated with the desired policy configuration. This file also supports adding other policies to be applied along the one being developed.

The playground/config directory can also contain other resource definitions, such as accessory services used by the policy (Eg. a remote authentication service).

For more information about make run, see Debugging Custom Policies Locally with PDK.

Test

The make test goal runs unit tests and integration tests. Integration tests are placed in the tests directory and are configured with the files placed at the tests/<module-name>/<test-name> directory.

For more information about writing integration tests, see Writing Integration Tests.

Publish

The make publish goal publishes the policy asset in Anypoint Exchange, in your configured Organization.

Since the publish goal is intended to publish a policy asset in development, the assetId and name published will explicitly say dev, and the versions published will include a timestamp at the end of the version. Eg.

groupId: your configured organization id
visible name: {Your policy name} Dev
assetId: {your-policy-asset-id}-dev
version: {your-policy-version}-20230618115723

For more information about publishing policies, see Uploading Custom Policies to Exchange.

Release

The make release goal also publishes the policy to Anypoint Exchange, but as a ready for production asset. In this case, the groupId, visible name, assetId and version will be the ones defined in the project.

For more information about releasing policies, see Uploading Custom Policies to Exchange.

Policy Examples

The PDK provides provides a set of example policy projects to get started creating policies and using the PDK features. To learn more about these examples see Custom policy Examples.

deejay-hub/openai-data-loss-prevention