Security, Compliance & Governance

Shared responsibility for AI on AWS

AWS secures the cloud (facilities, hypervisor, regional services). Customers secure in the cloud: IAM policies, encryption choices, VPC design, data classification, application prompts, and model configuration.

flowchart TB
  AWS[AWS: infrastructure platform compliance] --- CUST[Customer: data IAM app config prompts]
        

Definitions

Defense in depth

Layer controls (identity, network, encryption, logging) so no single failure exposes the system.

IAM for AI systems

Use least privilege roles for training jobs, inference endpoints, Lambda callers of Bedrock, and human operators. Separate data-access policies from model-invoke policies where possible.

Definitions

Service role

An IAM role an AWS service assumes to access S3, KMS keys, or other resources on your behalf.

Resource-based policy

Attached to resources (e.g. some model endpoints, buckets) defining who may access them—pairs with identity policies.

Encryption · Macie · PrivateLink

• AWS KMS — customer-managed keys for sensitive training and artifact stores
• Encryption at rest/in transit — defaults and mandatory TLS for APIs
• Amazon Macie — discover and protect sensitive data (PII) in S3-scale repositories
• AWS PrivateLink — private connectivity to AWS services without public internet exposure

Definitions

Data lineage

Traceability from raw datasets through features to model versions—supports audits and incident response.

Model Cards · catalogs · citations

SageMaker Model Cards and enterprise data catalogs document provenance, intended use, and evaluation results. GenAI apps should cite retrieved sources when claiming facts.

Definitions

Source citation

Linking generated statements to retrieval chunks or policy documents—reduces unsupported claims.

Data quality · privacy · integrity

Assess data quality, enforce access control on feature stores and vector indexes, and guard against poisoned uploads into RAG corpora.

Definitions

Prompt injection

Untrusted input that manipulates model behavior—mitigate with isolation, tooling boundaries, and monitoring.

Threat detection & monitoring

Combine CloudTrail (API audit), Amazon Inspector (workload vulnerabilities, where used), AWS Config (configuration compliance rules), and service-native logs for Bedrock/SageMaker to detect misuse.

Standards and assurance

Organizations map AI systems to programs like ISO family controls and SOC reports. Some jurisdictions emphasize algorithmic accountability—document decisions, testing, and oversight (high-level exam concept).

Definitions

AWS Artifact

Portal for on-demand compliance reports and agreements from AWS.

Audit Manager · Trusted Advisor

AWS Audit Manager helps collect evidence for audits continuously. AWS Trusted Advisor surfaces cost, fault tolerance, and security best-practice checks—useful hygiene, not a substitute for full AI risk review.

Lifecycle · residency · retention

Define policies for where data lives (Regions), how long it is retained, who can access embeddings and fine-tune datasets, and mandatory logging for sensitive invocations.

flowchart LR
  POL[Policies and standards] --> REV[Periodic review cadence]
  REV --> LOG[Logging monitoring evidence]
        

Frameworks such as the Generative AI Security Scoping Matrix (AWS discussion materials) help teams structure reviews across data, model, and application layers.

Governance cadence

Governance is not only tools: policies, review boards, training for builders and operators, and clear escalation for model incidents.

Domain 5 glossary

Shared responsibility

AWS vs customer security duties.

IAM · KMS · TLS

Access; keys; transport encryption.

Macie · PrivateLink

Sensitive data discovery; private service access.

CloudTrail · Config · Audit Manager · Artifact

Audit logs; drift rules; evidence automation; compliance downloads.

Lineage · Model Card · retention

Provenance docs; regulated lifecycles.

Prompt injection

App-layer threat for LLM systems.

You finished all five domains

• List three customer responsibilities under the shared responsibility model for AI
• Match Macie vs PrivateLink vs KMS to a business problem
• Name two services that support audit evidence collection
• Explain why vector/RAG data stores need access control and logging
• Define data residency vs retention in one line each

Use the separate 50-question practice exam HTML in this folder for mixed review.