FAQ
Q. What additional security layers should I consider to prevent prompt injection attacks?
A. The following diagram shows the three main security layers: LLM input, LLM built-in guardrails, and user-introduced guardrails.
Your organization should consider implementing security protocols across all layers. For
the first layer (LLM input), consider risk mitigation steps to help
secure the application by implementing mechanisms such as personally identifiable
information (PII) or sensitive information redaction, authentication, authorization, and
encryption. The second layer (LLM built-in guardrails) are model or
application securities provided by the LLM. Although most LLMs are trained with security
protocols to prevent inappropriate use, your organization should still consider adding
additional security controls by using Guardrails for Amazon Bedrock
Q. How can organizations defend against prompt injection attacks in prompt engineering?
A. Organizations can defend against prompt injection attacks by implementing best prompt engineering practices as discussed in the Best practices section. Your organization can also consider adding guardrails such as input validation, prompt sanitization, and secure communication channels.
Q. Are prompt security elements model-agnostic?
A. Generally, prompt security elements are designed for specific LLMs. Each LLM is trained differently in terms of data quality, diversity, representation, bias, and fine-tuning approaches, so a prompt security element that was introduced for one LLM isn't directly transferrable to another LLM. However, the security elements discussed in this guide can provide a framework and direction for developing tailored prompt security elements for other LLMs.
Q. How should I integrate these elements into an enterprise MLOps framework?
A. Depending on your organization's constraints and data
landscape, prompt security elements can be owned by the data scientist or developer who is
working on a specific generative AI use case or by a central generative AI governance team.
When you design the MLOps framework for a generative AI solution and release the solution to
the production environment, we recommend that you review the AWS blog posts FMOps/LLMOps: Operationalize generative AI and differences with MLOps
Q. What are some of the successful use cases?
A. The guardrails that are discussed in this guide were used successfully in RAG-based solutions for HR, corporate policy, insurance document summarization, corporate investment, and medical record summarization.
