Gadjit (pronounced "gadget") is an LLM-powered security bot framework designed to automate analyizing and taking action on access requests. Developed and open-sourced by Instacart, Gadjit leverages the power of Large Language Models (LLMs) to sift through identity information and seeks to end the status quo of "rubber stamp" access approvals commonly seen within manual access request processes.
Gadjit attempts to be a vendor-neutral framework for reviewing access requests. We achieve this with a plugin system which supports three different types of plugins:
- IGA Plugins: integrate with governance tools such as ConductorOne, Opal, or Lumos.
- LLM Plugins: integrate with LLMs such as OpenAI, Anthropic, or Gemini.
- Scoring Plugins: evaluate an access request based on a set of factors and return a score.
We strongly encourage contribution of new plugins, please open a pull request!
- OpenAI GPT-4o
- OpenAI GPT-4o via AWS API Gateway with IAM Authentication (Instacart uses an OpenAI proxy internally)
Analyzes the requested entitlement's members looking for commonality with the requester ("Does the requester's peers use this role in their day to day work?").
- Python 3.12
- Library dependencies (see
requirements.txt)
Gadjit can be run as a web server / AWS Lambda (best used for receiving webhook events), or as a one-shot application invoked with a cron job once per minute.
Package the Lambda zip. For an ARM-based Mac, this can be done with Docker:
docker run --platform linux/amd64 --mount type=bind,source=$PWD,target=/tmp --entrypoint "/bin/bash" -it public.ecr.aws/lambda/python:3.12 "-c" "dnf install -y zip && cd /tmp && rm -f lambda.zip && mkdir -p build && pip install . -t build && pip install -r requirements.txt -t build && cd build && zip -r ../lambda.zip * && cd .. && rm -rf build"
Deploy the resulting lambda.zip to AWS Lambda with at least 512MB of memory and at least a 120-second timeout. Set the runtime handler to gadjit.lambda_handler. Enable "Function URL" for your Lambda (with Auth Type=NONE and save. It is now ready to receive webhooks.
Install requirements:
pip install -Ur requirements.txt
Invoke:
python -m gadjit --server (--port 8080) (--config config.yaml)
Install requirements:
pip install -Ur requirements.txt
Invoke:
python -m gadjit (--config config.yaml)
Copy config.yaml.template to config.yaml and set the necessary values. A value can be defined in the YAML file or be referenced to an environmental variable by adding the prefix "env:"; whatever follows will be looked up in the current environment variables. For example:
gadjit:
log_level: "env:GADJIT_LOG_LEVEL"... means that the log level will be set to whatever the current value of the GADJIT_LOG_LEVEL env may be.
In some deployments (such as AWS Lambda), file-based configuration may not be ideal. In these cases, you can also configure Gadjit entirely with environment variables.
GADJIT__GADJIT__LOG_LEVEL=debug \
GADJIT__GADJIT__ENTITLEMENTS_TO_AUTO_APPROVE="Some-Entitlement-Name,2cxuJXhzhPoqvbJR75xq8ZVCBsK,Team - Security" \
GADJIT__IGA__PLUGINS__0__NAME=conductorone_cron \
GADJIT__IGA__PLUGINS__0__ENABLED=true \
GADJIT__IGA__PLUGINS__0__CONFIG__REASSIGN_TO_USER=2cxuJXhzhPoqvbJR75xq8ZVCBsK \
GADJIT__IGA__PLUGINS__0__CONFIG__BASE_URL="https://acme.conductor.one" \
GADJIT__IGA__PLUGINS__0__CONFIG__CLIENT_ID="strange-hydra-68836@acme.conductor.one/pcc" \
GADJIT__IGA__PLUGINS__0__CONFIG__CLIENT_SECRET="CONDUCTORONE_API_SECRET" \
GADJIT__SCORING__PLUGINS__0__NAME=requester_profile_attribute_proximity \
GADJIT__SCORING__PLUGINS__0__ENABLED=true \
GADJIT__LLM__PLUGINS__0__NAME=openai \
GADJIT__LLM__PLUGINS__0__ENABLED=true \
GADJIT__LLM__PLUGINS__0__CONFIG__SECRET_KEY="OPENAI_API_KEY"We welcome contributions from the community. We are especially interested in adding support for more IGA tools and additional configurability.
This project is licensed under the The 2.0 version of the Apache License - see the LICENSE file for details.
For questions or support, please open an issue in the repository.
Developed with ❤️ by Instacart