usnistgov/dioptra

Test Software for the Characterization of AI Technologies

Dioptra – A Test Platform for Trustworthy AI

What it is – Dioptra is an open‑source software platform (maintained by NIST) that lets you measure, analyze, and track the trustworthy characteristics of AI models. It implements the Measure function of the NIST AI Risk Management Framework, helping you evaluate safety, security, fairness, explainability, privacy, and other risk dimensions.

How it works – Dioptra runs as a set of Docker containers exposing a REST API. You can interact with it through:

  • an intuitive web UI,
  • a Python client library, or
  • any generic REST client. Experiments are defined in simple YAML files that compose modular components (e.g., data loaders, model runners, metric calculators). A plugin system lets you import existing Python packages, and a built‑in type system ensures interoperability between plugins.

Key properties

  • Reproducible – automatic snapshots of code, data, and configuration.
  • Traceable – full history of experiment inputs and results is stored.
  • Extensible – add new functionality via plugins.
  • Modular – build experiments from reusable YAML‑described blocks.
  • Secure & Multi‑tenant – user authentication (access controls forthcoming) and support for shared deployments.
  • Interactive – web UI for managing and visualising runs.

Typical use cases

Role Scenario
Model developer (1st‑party) Continuous testing of models during development to catch reliability or bias issues early.
Acquirer / evaluator (2nd‑party) Independent assessment of a vendor’s model before purchase or integration.
Auditor (3rd‑party) Formal compliance or regulatory audits of AI systems.
Researcher Tracking experiments for trustworthy‑AI studies, sharing results with collaborators.
Red‑team Running adversarial or stress‑test suites against a model in a controlled sandbox.

Getting started

  1. Pull the Docker images – core services (nginx, mlflow‑tracking, REST API) and worker images for PyTorch or TensorFlow (CPU/GPU).
  2. Create a deployment using the provided cookiecutter template (cruft create …).
  3. Initialize the deployment (./init-deployment.sh).
  4. Launch with docker compose up -d and open http://localhost.
  5. Follow the “Hello World” tutorial on the Dioptra docs site for a quick end‑to‑end run.

Documentation & community – Full user guides, installation steps, and tutorials are hosted at https://pages.nist.gov/dioptra/. Development guidelines are in DEVELOPER.md.

License – Creative Commons Attribution 4.0 International (CC‑BY‑4.0), a permissive open‑source license.

Citation – If you use Dioptra in a publication, cite the NIST DOI provided in the README.

Dioptra is a genuine, actively maintained project aimed at providing a reproducible, extensible framework for evaluating the trustworthiness of AI systems.

Related

  • Project
    dograh-hq/dograh

    Dograh AI is an open‑source, self‑hostable platform for building and testing voice conversational agents. It provides a visual workflow editor, telephony integrations (Twilio, Vonage, etc.), and a modular LLM/STT/TTS stack that you can replace with any provider. One‑command Docker setup gets the whole system running locally, and Python/Node SDKs let you create agents programmatically. The project is BSD‑2 licensed and backed by a Slack community.

  • Dispatch
    Anthropic Third-Party Testing AI Policy Proposal

    Anthropic proposes a third-party testing regime for frontier AI systems to validate safety, prevent national security risks, and avoid regulatory capture through a diverse ecosystem of auditors.

  • Project
    Vasco0x4/AIDA

    AIDA is an open‑source autonomous pentesting platform that lets any LLM (Claude, Gemini, OpenAI, etc.) drive a Docker‑based security‑testing environment. The agent can run reconnaissance, scanning, custom Python exploits, and raw HTTP manipulations, logging every command and its output in a persistent notebook. Findings are auto‑scored with CVSS 4.0 and can be exported as PDF reports. The project includes a web dashboard, JWT authentication, notification hooks, and supports swapping the built‑in `aida-pentest` container for the larger Exegol image. Real CVEs discovered by AIDA are listed, showing practical impact. Licensed under AGPL‑v3.

  • Project
    0xSteph/pentest-ai

    Pentest‑AI is an open‑source Python tool that lets you run LLM‑driven penetration‑testing against web applications. It generates probes via a connected AI model, then uses a deterministic oracle to reproduce each exploit and verify it against a control request. Verified findings are stored in replayable proof capsules and can be exported as SARIF for CI/CD pipelines. The package works with cloud models (Codex, Claude Code) or local Ollama models, is installed via `pipx`, and is MIT‑licensed.

  • Dispatch
    Mistral AI Shieldstral 1.0 3B Release

    Mistral AI has released Shieldstral, a 3B open-weights multimodal safety classifier that uses a binary question-answering framework to enable policy-adaptive moderation without retraining.