Title: AWS re:Inforce 2024 - Security controls for generative AI use cases (GAI221)

Insights:

The session focuses on security strategies for AI-specific applications, emphasizing vulnerability and threat modeling.
Adam Shostak's four questions for building a vulnerability matrix are highlighted: "What are we working on?", "What can go wrong?", and "What are we going to do about it?"
The scoping matrix is introduced to provide a common language for discussing AI security, ranging from consumer applications (Scope 1) to fully custom-built models (Scope 5).
Scope 1 and Scope 2 involve minimal changes to existing security practices, focusing on standard controls like role-based access controls (RBAC) and cloud access security brokers (CASB).
Scope 3 involves using pre-trained models and emphasizes the importance of retrieval augmented generation (RAG) for providing relevant information securely.
Scope 4 includes fine-tuning pre-trained models with domain-specific data, necessitating additional security measures to protect proprietary data and model artifacts.
Scope 5 involves creating custom models from scratch, requiring rigorous control over training data to avoid biases and ensure responsible AI practices.
Technical controls such as IAM, KMS, and Bedrock guardrails are essential for securing AI applications across different scopes.
Non-technical controls, including governance, processes, and training, are equally important for comprehensive AI security, as emphasized by the NIST Cybersecurity Framework (CSF) version 2.0.

"What we're talking about in terms of first principles today is vulnerability or threat modeling."
"What the matrix does is it helps you zero in on what type of application are we talking about specifically."
"In the wild, as it were, we have not seen really a host of new vulnerabilities."
"If something has made its way into the model weights, it can and will come out."
"Use your application to enforce those boundaries using the retrieval stage."
"You need to make sure that you protect those fine-tuned artifacts and the model inference endpoint."
"Not all controls are going to be technical."
"NIST CSF actually introduced an entirely new function called govern, which is all about processes, people, training, basically preparing your organization to adopt technologies in a safe manner."