c2.14 Protect data in generative-AI workloads

What "protect data in Gen AI" means

For a generative-AI workload, protecting data has a concrete meaning you can check: the data and prompts you send to a model (1) are not used by any provider to train their models, (2) are encrypted in transit and at rest, (3) travel over a private network path rather than the public internet, and (4) are filtered on the way in and out so sensitive data and prompt-injection attempts are caught. Two risks drive this: attackers trying to extract or manipulate data (for example prompt injection), and providers potentially using your prompts or fine-tuning data — provider terms and conditions may not be enough assurance.

How Amazon Bedrock delivers each control

• No training on your data — Amazon Bedrock does not use your prompts or fine-tuning data to train the base models, and no model provider can access them.
• Encryption — Bedrock requires encryption in transit and lets you encrypt fine-tuning data at rest with your own AWS KMS customer managed key.
• Private access — reach Bedrock from your VPC through a VPC endpoint without crossing the internet, and privately from on-premises over AWS Direct Connect.
• Input/output filtering — Amazon Bedrock Guardrails apply content filters, denied topics, and sensitive-information filters that block or mask PII in both prompts and responses, and detect prompt-injection attempts.

What alignment looks like

You are aligned with this control when, for every generative-AI workload:

• All foundation-model access goes through Amazon Bedrock (or a provider with equivalent, contractually guaranteed data-privacy and no-training terms you have verified).
• Fine-tuning and prompt data are encrypted at rest with a customer managed KMS key, and access to the model is over a private path (VPC endpoint / Direct Connect).
• Guardrails are configured and enforced, including sensitive-information (PII) filtering on inputs and outputs and prompt-attack protection.
• The workload appears in your AI inventory with an owner, and its data flows have been reviewed against your data-classification policy.

Go deeper with the AI Security Maturity Model

This is a foundational Cloud control covering the data-protection basics for Gen AI. If AI is central to your organization, the dedicated AI Security Maturity Model takes Gen-AI and agentic-AI data protection through the full maturity journey — use it as the deep-dive companion.

Related

Pair this with AI security governance for the inventory and ownership side, and encrypt data at rest for the KMS foundations.