IBM Introduces Framework for Securing Generative AI

The Growing Challenge of AI Security

As organizations continue to harness the power of generative AI, ensuring security has become a pressing concern. To address this challenge, IBM has unveiled a new security framework designed to help customers address the unique risks associated with generative AI workflows.

A Comprehensive Approach to Security

The IBM Framework for Securing Generative AI focuses on safeguarding gen AI workflows throughout the entire lifecycle, from data collection to production deployment. It offers guidance on the most likely security threats organizations will encounter when working with generative AI and provides recommendations for implementing effective defensive measures.

“We took our expertise and distilled it down to detail the most likely attacks along with the top defensive approaches that we think are the most important for organizations to focus on and to implement in order to secure their generative AI initiatives,” said Ryan Dougherty, program director, emerging security technology at IBM Security.

This framework is built on three core tenets: securing the data, the model, and the usage. Additionally, it emphasizes the importance of maintaining secure infrastructure and implementing AI governance throughout the process.

Unique Risks in Generative AI

Sridhar Muppidi, IBM Fellow and CTO at IBM Security, explained that while core data security practices like access control and infrastructure security remain crucial, generative AI comes with its own set of risks. These unique risks include:

• Data poisoning: The addition of false data to a dataset, leading to inaccurate results
• Bias and data diversity: Risks associated with biased or limited datasets in generative AI
• Data drift and data privacy: Ensuring the integrity and privacy of data throughout the generative AI process
• Prompt injection: The malicious modification of model output via a prompt

To address these risks, the IBM Framework for Securing Generative AI provides guidelines and suggestions for a range of tools and practices. It encompasses various security categories, including Machine Learning Detection and Response (MLDR), AI Security Posture Management (AISPM), and Machine Learning Security Operation (MLSecOps).

MLDR involves scanning models to identify potential risks, while AISPM focuses on ensuring secure deployments through appropriate configuration and best practices. “Just like we have DevOps and we added security and call DevSecOps, the idea is that MLSecOps is a whole end-to-end lifecycle, all the way from design to usage, and it provides that infusion of security,” explained Muppidi.