databricksruntime/blackice repository overview

BlackIce is a ready to run Docker image designed for AI security and red teaming. Inspired by Kali Linux, it addresses practical challenges faced by AI red teamers, including complex tool setup, dependency conflicts, isolated runtimes, and a fragmented tooling ecosystem.

The image bundles 14 version pinned open source tools spanning Responsible AI, security testing, and adversarial machine learning. Tools are exposed through a unified command line interface and can be run from the shell or within Databricks notebooks using a compute environment built from the image via Databricks Container Services.⁠

BlackIce separates tools into static and dynamic categories. Static tools are installed in isolated environments and require minimal programming, while dynamic tools support Python based customization for developing custom attacks. Dependency conflicts are explicitly managed within the container, allowing tools to coexist reliably. The image also includes minor patches that enable seamless interaction with Databricks Model Serving endpoints out of the box.

Additional documentation, examples, and Docker build artifacts are available in our GitHub Repo.⁠ Further details on the image architecture, tool selection process, and mapping to established AI risk frameworks can be found in the accompanying CAMLIS Red Paper.⁠