Title: AWS re:Inforce 2024 - Accelerate securely: The Generative AI Security Scoping Matrix (APS201)

Insights:

• Introduction and Purpose: The session was led by Matt Saner and Mike Lapidakis, who discussed the Generative AI Security Scoping Matrix, a tool designed to help organizations secure generative AI applications.
• Key Questions Addressed:
  • Where to start with generative AI security.
  • AWS's approach to security in generative AI.
  • Mechanisms for deploying generative AI securely.
  • Examples and applications of the security scoping matrix.
• Generative AI Overview: Generative AI is distinct from traditional AI/ML as it creates content (text, voice, video) and is non-deterministic. It democratizes AI, making it accessible to non-data scientists.
• Security Fundamentals: Emphasized the importance of "eating your security vegetables"—applying basic security principles to generative AI workloads.
• Three Legs of Generative AI Security:
  • Securing generative AI applications.
  • Leveraging security AI to enhance existing practices.
  • Protecting against novel threats from generative AI.
• Scopes and Domains: The matrix includes five scopes (consumer off-the-shelf applications, enterprise applications, pre-trained models, fine-tuned models, self-trained models) and five security domains (governance and compliance, legal and privacy, risk management, controls, resilience).
• Universal Truths: Avoid outright banning generative AI; instead, educate and empower users. Understand model versioning and emerging regulations.
• Scope-Specific Insights:
  • Scope 1: Consumer apps should not use PII or sensitive data. Focus on education and acceptable use policies.
  • Scope 2: Enterprise apps can use sensitive data due to legal agreements. Identity context and SLAs are crucial.
  • Scope 3: Building applications with foundation models. Emphasize threat modeling and understanding model training data.
  • Scope 4: Fine-tuning models with organizational data. Control access and avoid using sensitive information for fine-tuning.
  • Scope 5: Training models from scratch. Rare but requires comprehensive control over the entire process.
• Resilience and Controls: Each scope has specific resilience and control measures, such as identity management, encryption, and monitoring.
• Final Takeaways: The security scoping matrix is a flexible tool that can be adapted to organizational needs. AWS continues to develop and share best practices and documentation.

Quotes:

• "Generative AI is transforming our industry... it has democratized AI and also folks like Mike and myself and you, who might not be data scientists, can now take advantage of these radically transformative technologies."
• "We want to think about eating our security vegetables... get those basics right, apply those to the nuances of generative AI and understand those nuances."
• "The key here is to scope the workload, scope the conversation, scope the risks, and scope the controls."
• "Don't outright ban Gen AI. What you do want to do is empower your users, and you want to empower users through education, policy, controls on the back end that help enforce that, but also enablement."
• "If you are training or fine-tuning a model, the resiliency for the data that you might be using to do that drastically changes."
• "Scope three is where things fundamentally change with the generative AI security scoping matrix... your organization now owns the application, the security of that application."
• "In the current generation of LLMs, you can't think of it in the traditional relational database model... you'd have to retrain or refine tune the model."
• "Start with the security scoping matrix. Don't be afraid to use this. Adapt it to your organization's needs."