AIF-C01: Data Security Layer

AWS Audit Manager

Your Company's Evidence; continuous compliance auditing; automatically collects evidence to assess risk against frameworks (SOC 2 / HIPAA)

AWS Secrets Manager

Credential & API key management; securely stores; rotates; and retrieves database passwords and API keys automatically

VPC Interface Endpoint (PrivateLink)

Private ENI connection for Bedrock / SageMaker / & Rekognition APIs; keeps traffic off the public internet

Amazon EMR

Big Data Processing; distributed frameworks (PySpark / Hadoop); petabyte-scale data prep

AWS IAM (Identity & Access Management)

Authentication & authorization; fine-grained access control using Roles; Policies; and Least Privilege

VPC Interface Endpoint (PrivateLink)

Private ENI connection for Bedrock, SageMaker, & Rekognition APIs; keeps traffic off the public internet

Amazon EMR

Big Data Processing; distributed frameworks (PySpark, Hadoop); petabyte-scale data prep

AWS Data Exchange

Third-Party Datasets; independent software vendor (ISV) data feeds for ML training

AWS IAM (Identity & Access Management)

Authentication & authorization; fine-grained access control using Roles, Policies, and Least Privilege

AWS RAM (Resource Access Manager)

Cross-account resource sharing; securely share subnets or Transit Gateways across AWS Organizations

AWS Security Hub

Centralized security dashboard; aggregates alerts from multiple services; checks against CIS benchmarks

AWS Network Firewall

Managed network perimeter security; stateful inspection and web filtering for VPC traffic

Amazon Macie

Data privacy & PII discovery; uses ML and pattern matching to discover and protect sensitive data in Amazon S3

AWS CloudTrail

Governance & Compliance Audit Trail; logs and tracks all API actions/calls made across your AWS account

Prompt Injection

Malicious Inputs; user inputs designed to alter model behavior in unintended ways or bypass safety guardrails

Insecure Output Handling

Unsanitized Output; failure to clean or validate model outputs before passing them to downstream systems / users

Training Data Poisoning

Manipulated Training Sets; inserting harmful or biased data into training datasets to corrupt model behavior

Model Denial of Service (DoS)

Resource Exhaustion; exploiting vulnerabilities or high-cost prompts to render the model unavailable or unusable

Supply Chain Vulnerabilities

Third-Party Risks; security weaknesses in pre-trained models; datasets; or third-party components used in deployment

Sensitive Information Disclosure

Data Leakage; unauthorized exposure of confidential data or PII through model outputs or unintended channels

Insecure Plugin Design

Flawed Integrations; security flaws in optional extensions / plugins that can be exploited by malicious inputs

Excessive Agency

Unconstrained Autonomy; granting too much operational freedom or permissions to a model without human validation

Overreliance

Blind Trust; trusting model outputs excessively without adequate human review; validation; or verification

Model Theft

Unauthorized Exfiltration; unauthorized access; copying; or duplication of proprietary model parameters and weights