Post Syndicated from Matt Saner original https://aws.amazon.com/blogs/security/securing-generative-ai-an-introduction-to-the-generative-ai-security-scoping-matrix/

Generative artificial intelligence (generative AI) has captured the imagination of organizations and is transforming the customer experience in industries of every size across the globe. This leap in AI capability, fueled by multi-billion-parameter large language models (LLMs) and transformer neural networks, has opened the door to new productivity improvements, creative capabilities, and more.

As organizations evaluate and adopt generative AI for their employees and customers, cybersecurity practitioners must assess the risks, governance, and controls for this evolving technology at a rapid pace. As security leaders working with the largest, most complex customers at Amazon Web Services (AWS), we’re regularly consulted on trends, best practices, and the rapidly evolving landscape of generative AI and the associated security and privacy implications. In that spirit, we’d like to share key strategies that you can use to accelerate your own generative AI security journey.

This post, the first in a series on securing generative AI, establishes a mental model that will help you approach the risk and security implications based on the type of generative AI workload you are deploying. We then highlight key considerations for security leaders and practitioners to prioritize when securing generative AI workloads. Follow-on posts will dive deep into developing generative AI solutions that meet customers’ security requirements, best practices for threat modeling generative AI applications, approaches for evaluating compliance and privacy considerations, and will explore ways to use generative AI to improve your own cybersecurity operations.

Where to start

As with any emerging technology, a strong grounding in the foundations of that technology is critical to helping you understand the associated scopes, risks, security, and compliance requirements. To learn more about the foundations of generative AI, we recommend that you start by reading more about what generative AI is, its unique terminologies and nuances, and exploring examples of how organizations are using it to innovate for their customers.

If you’re just starting to explore or adopt generative AI, you might imagine that an entirely new security discipline will be required. While there are unique security considerations, the good news is that generative AI workloads are, at their core, another data-driven computing workload, and they inherit much of the same security regimen. The fact is, if you’ve invested in cloud cybersecurity best practices over the years and embraced prescriptive advice from sources like Steve’s top 10, the Security Pillar of the Well-Architected Framework, and the Well-Architected Machine Learning Lens, you’re well on your way!

Core security disciplines, like identity and access management, data protection, privacy and compliance, application security, and threat modeling are still critically important for generative AI workloads, just as they are for any other workload. For example, if your generative AI application is accessing a database, you’ll need to know what the data classification of the database is, how to protect that data, how to monitor for threats, and how to manage access. But beyond emphasizing long-standing security practices, it’s crucial to understand the unique risks and additional security considerations that generative AI workloads bring. This post highlights several security factors, both new and familiar, for you to consider.

With that in mind, let’s discuss the first step: scoping.

Determine your scope

Your organization has made the decision to move forward with a generative AI solution; now what do you do as a security leader or practitioner? As with any security effort, you must understand the scope of what you’re tasked with securing. Depending on your use case, you might choose a managed service where the service provider takes more responsibility for the management of the service and model, or you might choose to build your own service and model.

Let’s look at how you might use various generative AI solutions in the AWS Cloud. At AWS, security is a top priority, and we believe providing customers with the right tool for the job is critical. For example, you can use the serverless, API-driven Amazon Bedrock with simple-to-consume, pre-trained foundation models (FMs) provided by AI21 Labs, Anthropic, Cohere, Meta, stability.ai, and Amazon Titan. Amazon SageMaker JumpStart provides you with additional flexibility while still using pre-trained FMs, helping you to accelerate your AI journey securely. You can also build and train your own models on Amazon SageMaker. Maybe you plan to use a consumer generative AI application through a web interface or API such as a chatbot or generative AI features embedded into a commercial enterprise application your organization has procured. Each of these service offerings has different infrastructure, software, access, and data models and, as such, will result in different security considerations. To establish consistency, we’ve grouped these service offerings into logical categorizations, which we’ve named scopes.

In order to help simplify your security scoping efforts, we’ve created a matrix that conveniently summarizes key security disciplines that you should consider, depending on which generative AI solution you select. We call this the Generative AI Security Scoping Matrix, shown in Figure 1.

Figure 1: Generative AI Security Scoping Matrix

The first step is to determine which scope your use case fits into. The scopes are numbered 1–5, representing least ownership to greatest ownership.

Buying generative AI:

• Scope 1: Consumer app – Your business consumes a public third-party generative AI service, either at no-cost or paid. At this scope you don’t own or see the training data or the model, and you cannot modify or augment it. You invoke APIs or directly use the application according to the terms of service of the provider.
  Example: An employee interacts with a generative AI chat application to generate ideas for an upcoming marketing campaign.
• Scope 2: Enterprise app – Your business uses a third-party enterprise application that has generative AI features embedded within, and a business relationship is established between your organization and the vendor.
  Example: You use a third-party enterprise scheduling application that has a generative AI capability embedded within to help draft meeting agendas.

Building generative AI:

• Scope 3: Pre-trained models – Your business builds its own application using an existing third-party generative AI foundation model. You directly integrate it with your workload through an application programming interface (API).
  Example: You build an application to create a customer support chatbot that uses the Anthropic Claude foundation model through Amazon Bedrock APIs.
• Scope 4: Fine-tuned models – Your business refines an existing third-party generative AI foundation model by fine-tuning it with data specific to your business, generating a new, enhanced model that’s specialized to your workload.
  Example: Using an API to access a foundation model, you build an application for your marketing teams that enables them to build marketing materials that are specific to your products and services.
• Scope 5: Self-trained models – Your business builds and trains a generative AI model from scratch using data that you own or acquire. You own every aspect of the model.
  Example: Your business wants to create a model trained exclusively on deep, industry-specific data to license to companies in that industry, creating a completely novel LLM.

In the Generative AI Security Scoping Matrix, we identify five security disciplines that span the different types of generative AI solutions. The unique requirements of each security discipline can vary depending on the scope of the generative AI application. By determining which generative AI scope is being deployed, security teams can quickly prioritize focus and assess the scope of each security discipline.

Let’s explore each security discipline and consider how scoping affects security requirements.

• Governance and compliance – The policies, procedures, and reporting needed to empower the business while minimizing risk.
• Legal and privacy – The specific regulatory, legal, and privacy requirements for using or creating generative AI solutions.
• Risk management – Identification of potential threats to generative AI solutions and recommended mitigations.
• Controls – The implementation of security controls that are used to mitigate risk.
• Resilience – How to architect generative AI solutions to maintain availability and meet business SLAs.

Throughout our Securing Generative AI blog series, we’ll be referring to the Generative AI Security Scoping Matrix to help you understand how various security requirements and recommendations can change depending on the scope of your AI deployment. We encourage you to adopt and reference the Generative AI Security Scoping Matrix in your own internal processes, such as procurement, evaluation, and security architecture scoping.

What to prioritize

Your workload is scoped and now you need to enable your business to move forward fast, yet securely. Let’s explore a few examples of opportunities you should prioritize.

Governance and compliance plus Legal and privacy

Figure 2: Governance and compliance

With consumer off-the-shelf apps (Scope 1) and enterprise off-the-shelf apps (Scope 2), you must pay special attention to the terms of service, licensing, data sovereignty, and other legal disclosures. Outline important considerations regarding your organization’s data management requirements, and if your organization has legal and procurement departments, be sure to work closely with them. Assess how these requirements apply to a Scope 1 or 2 application. Data governance is critical, and an existing strong data governance strategy can be leveraged and extended to generative AI workloads. Outline your organization’s risk appetite and the security posture you want to achieve for Scope 1 and 2 applications and implement policies that specify that only appropriate data types and data classifications should be used. For example, you might choose to create a policy that prohibits the use of personal identifiable information (PII), confidential, or proprietary data when using Scope 1 applications.

If a third-party model has all the data and functionality that you need, Scope 1 and Scope 2 applications might fit your requirements. However, if it’s important to summarize, correlate, and parse through your own business data, generate new insights, or automate repetitive tasks, you’ll need to deploy an application from Scope 3, 4, or 5. For example, your organization might choose to use a pre-trained model (Scope 3). Maybe you want to take it a step further and create a version of a third-party model such as Amazon Titan with your organization’s data included, known as fine-tuning (Scope 4). Or you might create an entirely new first-party model from scratch, trained with data you supply (Scope 5).

In Scopes 3, 4, and 5, your data can be used in the training or fine-tuning of the model, or as part of the output. You must understand the data classification and data type of the assets the solution will have access to. Scope 3 solutions might use a filtering mechanism on data provided through Retrieval Augmented Generation (RAG) with the help from Agents for Amazon Bedrock, for example, as an input to a prompt. RAG offers you an alternative to training or fine-tuning by querying your data as part of the prompt. This then augments the context for the LLM to provide a completion and response that can use your business data as part of the response, rather than directly embedding your data in the model itself through fine-tuning or training. See Figure 3 for an example data flow diagram demonstrating how customer data could be used in a generative AI prompt and response through RAG.

Figure 3: Retrieval Augmented Generation (RAG)

In scopes 4 and 5, on the other hand, you must classify the modified model for the most sensitive level of data classification used to fine-tune or train the model. Your model would then mirror the data classification on the data it was trained against. For example, if you supply PII in the fine-tuning or training of a model, then the new model will contain PII. Currently, there are no mechanisms for easily filtering the model’s output based on authorization, and a user could potentially retrieve data they wouldn’t otherwise be authorized to see. Consider this a key takeaway; your application can be built around your model to implement filtering controls on your business data as part of a RAG data flow, which can provide additional data security granularity without placing your sensitive data directly within the model.

Figure 4: Legal and privacy

From a legal perspective, it’s important to understand both the service provider’s end-user license agreement (EULA), terms of services (TOS), and any other contractual agreements necessary to use their service across Scopes 1 through 4. For Scope 5, your legal teams should provide their own contractual terms of service for any external use of your models. Also, for Scope 3 and Scope 4, be sure to validate both the service provider’s legal terms for the use of their service, as well as the model provider’s legal terms for the use of their model within that service.

Additionally, consider the privacy concerns if the European Union’s General Data Protection Regulation (GDPR) “right to erasure” or “right to be forgotten” requirements are applicable to your business. Carefully consider the impact of training or fine-tuning your models with data that you might need to delete upon request. The only fully effective way to remove data from a model is to delete the data from the training set and train a new version of the model. This isn’t practical when the data deletion is a fraction of the total training data and can be very costly depending on the size of your model.

Risk management

Figure 5: Risk management

While AI-enabled applications can act, look, and feel like non-AI-enabled applications, the free-form nature of interacting with an LLM mandates additional scrutiny and guardrails. It is important to identify what risks apply to your generative AI workloads, and how to begin to mitigate them.

There are many ways to identify risks, but two common mechanisms are risk assessments and threat modeling. For Scopes 1 and 2, you’re assessing the risk of the third-party providers to understand the risks that might originate in their service, and how they mitigate or manage the risks they’re responsible for. Likewise, you must understand what your risk management responsibilities are as a consumer of that service.

For Scopes 3, 4, and 5—implement threat modeling—while we will dive deep into specific threats and how to threat-model generative AI applications in a future blog post, let’s give an example of a threat unique to LLMs. Threat actors might use a technique such as prompt injection: a carefully crafted input that causes an LLM to respond in unexpected or undesired ways. This threat can be used to extract features (features are characteristics or properties of data used to train a machine learning (ML) model), defame, gain access to internal systems, and more. In recent months, NIST, MITRE, and OWASP have published guidance for securing AI and LLM solutions. In both the MITRE and OWASP published approaches, prompt injection (model evasion) is the first threat listed. Prompt injection threats might sound new, but will be familiar to many cybersecurity professionals. It’s essentially an evolution of injection attacks, such as SQL injection, JSON or XML injection, or command-line injection, that many practitioners are accustomed to addressing.

Emerging threat vectors for generative AI workloads create a new frontier for threat modeling and overall risk management practices. As mentioned, your existing cybersecurity practices will apply here as well, but you must adapt to account for unique threats in this space. Partnering deeply with development teams and other key stakeholders who are creating generative AI applications within your organization will be required to understand the nuances, adequately model the threats, and define best practices.

Controls

Figure 6: Controls

Controls help us enforce compliance, policy, and security requirements in order to mitigate risk. Let’s dive into an example of a prioritized security control: identity and access management. To set some context, during inference (the process of a model generating an output, based on an input) first- or third-party foundation models (Scopes 3–5) are immutable. The API to a model accepts an input and returns an output. Models are versioned and, after release, are static. On its own, the model itself is incapable of storing new data, adjusting results over time, or incorporating external data sources directly. Without the intervention of data processing capabilities that reside outside of the model, the model will not store new data or mutate.

Both modern databases and foundation models have a notion of using the identity of the entity making a query. Traditional databases can have table-level, row-level, column-level, or even element-level security controls. Foundation models, on the other hand, don’t currently allow for fine-grained access to specific embeddings they might contain. In LLMs, embeddings are the mathematical representations created by the model during training to represent each object—such as words, sounds, and graphics—and help describe an object’s context and relationship to other objects. An entity is either permitted to access the full model and the inference it produces or nothing at all. It cannot restrict access at the level of specific embeddings in a vector database. In other words, with today’s technology, when you grant an entity access directly to a model, you are granting it permission to all the data that model was trained on. When accessed, information flows in two directions: prompts and contexts flow from the user through the application to the model, and a completion returns from the model back through the application providing an inference response to the user. When you authorize access to a model, you’re implicitly authorizing both of these data flows to occur, and either or both of these data flows might contain confidential data.

For example, imagine your business has built an application on top of Amazon Bedrock at Scope 4, where you’ve fine-tuned a foundation model, or Scope 5 where you’ve trained a model on your own business data. An AWS Identity and Access Management (IAM) policy grants your application permissions to invoke a specific model. The policy cannot limit access to subsets of data within the model. For IAM, when interacting with a model directly, you’re limited to model access.

{
	"Version": "2012-10-17",
	"Statement": {
		"Sid": "AllowInference",
		"Effect": "Allow",
		"Action": [
			"bedrock:InvokeModel"
		],
		"Resource": "arn:aws:bedrock:*::<foundation-model>/<model-id-of-model-to-allow>
	}
}

What could you do to implement least privilege in this case? In most scenarios, an application layer will invoke the Amazon Bedrock endpoint to interact with a model. This front-end application can use an identity solution, such as Amazon Cognito or AWS IAM Identity Center, to authenticate and authorize users, and limit specific actions and access to certain data accordingly based on roles, attributes, and user communities. For example, the application could select a model based on the authorization of the user. Or perhaps your application uses RAG by querying external data sources to provide just-in-time data for generative AI responses, using services such as Amazon Kendra or Amazon OpenSearch Serverless. In that case, you would use an authorization layer to filter access to specific content based on the role and entitlements of the user. As you can see, identity and access management principles are the same as any other application your organization develops, but you must account for the unique capabilities and architectural considerations of your generative AI workloads.

Resilience

Figure 7: Resilience

Finally, availability is a key component of security as called out in the C.I.A. triad. Building resilient applications is critical to meeting your organization’s availability and business continuity requirements. For Scope 1 and 2, you should understand how the provider’s availability aligns to your organization’s needs and expectations. Carefully consider how disruptions might impact your business should the underlying model, API, or presentation layer become unavailable. Additionally, consider how complex prompts and completions might impact usage quotas, or what billing impacts the application might have.

For Scopes 3, 4, and 5, make sure that you set appropriate timeouts to account for complex prompts and completions. You might also want to look at prompt input size for allocated character limits defined by your model. Also consider existing best practices for resilient designs such as backoff and retries and circuit breaker patterns to achieve the desired user experience. When using vector databases, having a high availability configuration and disaster recovery plan is recommended to be resilient against different failure modes.

Instance flexibility for both inference and training model pipelines are important architectural considerations in addition to potentially reserving or pre-provisioning compute for highly critical workloads. When using managed services like Amazon Bedrock or SageMaker, you must validate AWS Region availability and feature parity when implementing a multi-Region deployment strategy. Similarly, for multi-Region support of Scope 4 and 5 workloads, you must account for the availability of your fine-tuning or training data across Regions. If you use SageMaker to train a model in Scope 5, use checkpoints to save progress as you train your model. This will allow you to resume training from the last saved checkpoint if necessary.

Be sure to review and implement existing application resilience best practices established in the AWS Resilience Hub and within the Reliability Pillar and Operational Excellence Pillar of the Well Architected Framework.

Conclusion

In this post, we outlined how well-established cloud security principles provide a solid foundation for securing generative AI solutions. While you will use many existing security practices and patterns, you must also learn the fundamentals of generative AI and the unique threats and security considerations that must be addressed. Use the Generative AI Security Scoping Matrix to help determine the scope of your generative AI workloads and the associated security dimensions that apply. With your scope determined, you can then prioritize solving for your critical security requirements to enable the secure use of generative AI workloads by your business.

Please join us as we continue to explore these and additional security topics in our upcoming posts in the Securing Generative AI series.

If you have feedback about this post, submit comments in the Comments section below. If you have questions about this post, contact AWS Support.

Want more AWS Security news? Follow us on Twitter.