Domain 4: Secure AI Architecture, Engineering, and Identity and Access Management (IAM) (18% of exam)

0.0(0)
Studied by 0 people
call kaiCall Kai
learnLearn
examPractice Test
spaced repetitionSpaced Repetition
heart puzzleMatch
flashcardsFlashcards
GameKnowt Play
Card Sorting

1/32

encourage image

There's no tags or description

Looks like no tags are added yet.

33 Terms

1
New cards
AI-specific security strategies
Foundational design approaches applied to AI systems, including Zero Trust (verify everything, trust nothing by default), Defense-in-Depth (layered controls so no single failure is catastrophic), Privacy-by-Design (privacy built in from the start), and Secure Defaults (safest configuration out of the box rather than opt-in security).
2
New cards
AI threat modeling across architectural components
Applying threat modeling specifically to distinct AI attack surfaces, such as RAG pipelines (poisoned retrieval content), vector databases (embedding leakage), and inference endpoints (query-based model extraction).
3
New cards
Architectural trade-offs (AI)
Recognizing inherent design tensions when architecting AI systems—specifically balancing latency, explainability, and cost (for example, a more explainable model may be slower or more expensive to run).
4
New cards
Build vs. buy (AI solutions)
Evaluating hosted AI solutions (vendor-managed, faster to deploy, less control) versus self-managed AI solutions (greater control and customization, higher internal security and operational burden).
5
New cards
MCP (Model Context Protocol)
A standardized protocol for securely connecting AI models to external tools and data sources.
6
New cards
A2A (Agent-to-Agent) protocols
Standardized communication protocols enabling multiple autonomous AI agents to interact and coordinate securely with each other.
7
New cards
Hardening AI deployments
Applying standard infrastructure hardening practices across cloud, on-premises, and hybrid environments wherever AI workloads execute.
8
New cards
Isolation and sandboxing (AI infrastructure)
Separating training and inference workloads from each other and from adjacent systems so a compromise or failure in one environment does not cascade.
9
New cards
Availability controls against resource exhaustion
Implementing rate limiting and compute throttling on AI endpoints to prevent a single user or attacker from starving the system of GPU or compute resources.
10
New cards
Isolated accelerator capacity
Dedicating or segmenting GPU and TPU hardware resources so workloads—especially in multi-tenant environments—cannot interfere with or snoop on one another.
11
New cards
Physical and environmental security (AI)
Standard data center controls—including power, cooling, rack security, and fire suppression—emphasized because dense AI GPU clusters impose intense power and thermal demands.
12
New cards
Secure CI/CD and model deployment pipelines
Applying the same packaging, containerization, and security rigor to AI model deployment pipelines as traditional application code pipelines.
13
New cards
Integrity and confidentiality controls for data pipelines
Protecting training and inference data as it moves through ingestion and transformation pipeline stages, rather than securing data solely at rest.
14
New cards
Lineage tracking for fine-tuning
Maintaining verifiable records of the exact datasets and base models used to create a fine-tuned model, enabling audits and root-cause tracing when issues arise.
15
New cards
Validation checkpoints and pipeline gating
Automated quality and security gates (tests and approvals) that a model must pass before advancing to the next MLOps/LLMOps pipeline stage, preventing flawed models from reaching production.
16
New cards
Data unlearning capabilities
Technical mechanisms designed to make a trained model forget specific data it learned from, addressing regulatory right-to-erasure requirements.
17
New cards
AI input/output handling
Technical implementation of guardrails including prompt sanitization, content filtering, and context boundary validation to control what enters and exits a model.
18
New cards
Architectural guardrails for autonomous behavior
Structural wrappers—such as an agent harness (constraining what an agent can do) and orchestration (coordinating multi-step or multi-agent workflows within defined limits)—that bound autonomous AI actions.
19
New cards
Functional, performance, and load testing (AI)
Testing protocols that confirm an AI system operates accurately and maintains acceptable performance under both expected and peak traffic conditions.
20
New cards
Stress testing and exception handling validation (AI)
Testing protocols that verify an AI system fails safely—without taking dangerous or unauthorized actions—when pushed beyond normal limits or given unexpected inputs.
21
New cards
Human vs. non-human identity classification
Distinguishing between human user accounts and non-human identities—such as AI agents, workloads, services, and tools—so each receives its own distinct, manageable identity.
22
New cards
Identity architectures for non-repudiation and accountability
Ensuring every AI action can be traced back to a verifiable identity, establishing not just which agent acted, but under whose authorization it operated.
23
New cards
M2M (Machine-to-Machine) authentication and API tokens
Mechanisms enabling non-human AI identities, services, and agents to authenticate directly to one another using traditional IAM controls.
24
New cards
Federated identities, MFA, and SSO (AI)
Standard enterprise identity mechanisms extended across AI platforms to govern both human users and AI services where applicable.
25
New cards
Secrets and key management for AI platforms
Protecting the API keys, tokens, and credentials that AI systems and agents use to authenticate, carrying heightened risk due to autonomous agent execution.
26
New cards
PoLP and SoD across AI and agentic workflows
Applying Principle of Least Privilege (granting only necessary permissions) and Separation of Duties (preventing any single agent or user from controlling an entire sensitive process) to models and agents.
27
New cards
AI authorization methodologies (RBAC, ABAC, PBAC, JIT)
Access control models applied to AI: RBAC (tied to defined roles), ABAC (tied to dynamic attributes like time, location, or data sensitivity), PBAC (governed by organizational policy rules), and JIT (Just-in-Time elevated access granted only for a specific window and then revoked).
28
New cards
Execution boundaries and API scope limitations
Strictly restricting the specific actions and API calls an AI agent is permitted to invoke, even if broader underlying access technically exists.
29
New cards
Excessive agency, over-delegation, and lateral movement prevention
Mitigating the agentic AI risk where an overly autonomous or broadly scoped agent takes unintended actions or pivots beyond its intended task through autonomous decision-making.
30
New cards
AI identity inventories
Maintaining a complete, centralized registry of both human and non-human (agent and service) identities so all entities interacting with AI systems are visible and governed.
31
New cards
Identity provisioning and deprovisioning (AI)
Creating and fully revoking access credentials when an AI agent, service, or human user no longer requires access, preventing orphaned or lingering privileges.
32
New cards
Periodic access reviews and ISPM
Regularly re-validating that granted AI permissions remain appropriate, supported by Identity Security Posture Management (ISPM) tools for continuous access monitoring.
33
New cards
Audit of AI resource utilization
Monitoring token accounting, rate limits, and session timeouts as both a cost-control measure and a security control, since anomalous token consumption can indicate misuse or compromise.