This AI Ecosystem Report, featuring CISO Chris Hughes, an Acceleration Economy practitioner analyst, looks at Databricks’ AI Security Framework.
Highlights
00:09 — Databricks’ AI Security Framework starts off covering AI and machine learning (ML) model types: predictive ML models like PyTorch and Hugging Face; state-of-the-art open models like Llama, and external models of third-party services like OpenAI’s ChatGPT and Anthropic.
01:13 — The framework covers four system stages: data operations, model operations, model deployment and serving, and operations and platform. First: data operations. Risks include insufficient access controls, missing data classifications, or poor data quality. Next, model operations: Risks include model drift, ML supply chain vulnerabilities, and model theft.
02:30 — Next up is model deployment and serving. This includes components like model serving inference requests or responses. Some of the risks include prompt injection, model breakout, and output manipulation.
03:09 — The last system stage is called operations and platform. Some of the risks here will look very familiar because they’re broader cybersecurity risks. These include a lack of enforcement and repeatable standards as well as a lack of vulnerability management, compliance, and incident response.
04:02 — The Databricks platform, and how it addresses these risks, is covered in great detail, using specific examples from some of its customers.
