Domain 3: AI Data Management, Privacy, and Model Security (19% of exam)

Data ownership roles

Information owner is accountable for business purpose and risk, steward manages day-to-day quality and handling, and controller determines the purposes and means of processing under GDPR-style frameworks.

Data sovereignty and cross-border residency

Legal requirements stipulating that data, and sometimes the AI processing of it, must remain within specific national or regional borders.

Data classification and tagging

Labeling data by sensitivity levels (public, internal, confidential, restricted) so handling rules apply automatically, paired with sensitive data impact assessments to evaluate exposure or misuse consequences in AI systems.

Data provenance, lineage, and sourcing

Tracking where data originated and every transformation it underwent, including distinguishing between real and synthetic sourced data to mitigate model collapse risks.

Chain-of-custody

A documented, unbroken forensic record of who handled a dataset and when, ensuring evidentiary integrity — same forensic concept as physical evidence, applied to datasets.

AI data lifecycle

collection → validation → transformation → retention → archiving → deletion (an expanded version of the 1.4 lifecycle, now with governance/accountability framing).

Privacy by Design principles

Building privacy into AI systems from the start through consent, purpose limitation (using data only for stated reasons), lawful usage, and data minimization (collecting only what is necessary).

Data protection techniques

Methods to safeguard sensitive information in AI pipelines, including tokenization, anonymization, differential privacy (adding statistical noise to protect individual records while preserving aggregate patterns), masking, and noise injection.

Global privacy regulations

Legal frameworks governing specific data domains that apply to AI data handling, such as GDPR (EU general data protection), HIPAA (US health data), and PCI DSS (payment card data).

AI-specific privacy challenges

Risks unique to AI models, including unintentional bias (learning correlations that disadvantage groups without explicit instruction) and difficulty fulfilling the GDPR right to erasure once data is baked into trained model weights.

Training data integrity

Security controls ensuring that the data used to train an AI model has not been tampered with or corrupted.

Data poisoning, backdoors, and extraction threats

Mitigating attacks that corrupt training data, embed hidden triggers, or reconstruct training data from model outputs specifically at the data-handling stage.

Unauthorized label modification

A subtle form of data poisoning where an attacker alters labels in training data (such as mislabeling malware as safe) to corrupt what the model learns.

Malicious datasets

Entire datasets intentionally crafted or injected to harm a model, contrasted with legitimate datasets that are tampered with after creation.

Dataset tampering and versioning

Detecting unauthorized modifications to a dataset and maintaining version history so security teams can verify integrity and revert to a known-good state.

Recursive training data security

Protecting against model collapse by controlling how much AI-generated content re-enters future training sets and verifying synthetic data trustworthiness so it remains representative and uncorrupted.

DSPM (Data Security Posture Management)

Tools and practices for continuously discovering and assessing security risks across where sensitive data resides and how it flows, including into AI systems.

DLP (Data Loss Prevention) and semantic filtering

Combining traditional pattern-matching DLP (catching SSNs or credit cards) with semantic filtering on AI inputs and outputs to catch meaning-based leaks, such as paraphrased secrets.

Encryption for data states

Protecting data in use (actively being processed, the hardest state to secure), in transit (moving across networks), and at rest (stored) across AI pipelines and model weights.

Key management lifecycle

Securing cryptographic keys across AI environments using centralized KMS, BYOK (Bring Your Own Keys for customer control over vendor environments), secret management, and regular key rotation.

Purpose-based and restricted data access

Access controls tied to why data is being accessed rather than just user identity, such as an Agentic AI data catalog restricting an agent to data strictly relevant to its authorized task.

Media sanitization on high-density AI storage

Using cryptographic erasure (destroying encryption keys to render data unreadable) and micro-shredding to sanitize specialized drives and large GPU clusters that cannot be wiped with traditional methods.

Model governance

Establishing clear ownership, risk classification, and formal approval workflows for AI models throughout their lifecycle.

Model documentation and approved model registries

Using standardized model cards to document intended use, training data, limitations, and performance, alongside registries that track which models are approved for deployment.

Model deployment pipeline security

Securing validation, packaging, serving, and credential management as an AI model transitions from development into production environments.

Input and output sanitization, updates, and rollback

Continuous operational safeguards applied post-deployment to filter malicious prompts/responses, patch vulnerabilities, and revert to safe model versions when issues arise.

Security lifecycle monitoring

Ongoing post-deployment surveillance watching for model drift, bias degradation, and anomalous behavior.

Privacy-preserving audit trails

Logging AI data events in a manner that supports forensic investigation and accountability without exposing sensitive personal or proprietary data within the logs themselves.

Model confidentiality

Protecting proprietary model weights, algorithms, and intellectual property from unauthorized access or theft as critical organizational assets.

Model extraction prevention

Defenses designed to stop attackers from reverse-engineering or stealing a model's functionality and weights through repeated querying.

Watermarking and fingerprinting

Embedding identifiable markers into a model's outputs or weights to prove ownership and detect unauthorized theft or misuse.

Model integrity

Using model signing, cryptographic hashing, and strict version control to prove an AI model has not been altered since validation.

Model availability

Implementing rate limiting, quota enforcement, and graceful degradation to keep an AI model usable and responsive under heavy load or denial-of-service attacks without suffering a full outage.

Supply chain risk assessments

Evaluating security, privacy, and integrity risks introduced by third-party, open-source, and public datasets used during model training or fine-tuning.

Provenance and authenticity verification for vendors

Confirming that a third-party AI model or dataset genuinely originates from and matches the claims of its vendor (pedigree verification).

Foundation and pre-trained model risks

Inheriting unknown vulnerabilities, systemic biases, or hidden backdoors from third-party base models due to a lack of full visibility into how they were constructed and trained.