CISSP Domain 3: Security Architecture and Engineering Guide — Build Security In, Not On

Security Architecture and Engineering is where security stops being policy and becomes structure. Domain 3 is 13% of the CISSP exam and the broadest technically: it spans secure design principles (least privilege, defense in depth, zero trust, secure defaults), the formal security models that prove a system enforces confidentiality or integrity, cryptography from symmetric ciphers to NIST's new post-quantum standards, and the physical engineering of data centers, power, and fire suppression. In a real job this is the architect's domain, the person who decides how a control is enforced before the first line of code, so that a single misconfiguration cannot collapse the whole system. Get this right and security is built in; get it wrong and every later domain spends its budget compensating.

Domain 3 at a glance

Flip each card for the one-line essence of each area before you dive in.

Secure design principles

Think of a bank vault: one guard cannot open it alone, the door locks itself if power dies, and even insiders get only the keys their job needs. That is secure design in one image. The 2024 ISC2 outline frames these as secure design principles you bake in before a single line of code ships, not bolt on after a breach.

Least privilege grants every user, service, and process only the permissions its task demands and nothing more. A backup script needs read access, never domain-admin rights. Separation of duties splits a sensitive action so no one person can complete it alone. The developer who writes code should not also approve its push to production. Defense in depth layers independent controls so one failure never collapses the whole system: firewall, then segmentation, then MFA, then logging. Zero trust assumes no user or device is trusted, even inside the LAN, and re-verifies each request.

Secure defaults mean the product ships locked down: the safest configuration is the out-of-box state, so a lazy admin is still safe. Fail-secure means a crashed firewall blocks traffic instead of passing it. Threat modeling (STRIDE: Spoofing, Tampering, Repudiation, Information disclosure, DoS, Elevation of privilege) finds weaknesses on the whiteboard, before attackers find them in production. Together they form the secure-by-design mindset.

Security models

Think of a military mailroom: a clerk with "Secret" clearance can read a "Confidential" memo, but can never leak it into a "Top Secret" file. That intuition is Bell-LaPadula. Security models are formal rule-sets that map a security goal onto provable access rules. CISSP tests them through scenarios, so learn the goal each one protects.

Bell-LaPadula (BLP) protects confidentiality using two rules. The Simple Security Property is no read up: a subject cannot read an object at a higher classification. The Star (*) Property is no write down: a subject cannot write to a lower level, preventing leaks. A discretionary rule layers an access matrix on top.

Biba inverts BLP to protect integrity. The Simple Integrity Axiom is no read down (do not consume dirtier data). The Star Integrity Axiom is no write up (do not contaminate cleaner data). Memorise it as the mirror image of BLP.

Clark-Wilson protects integrity commercially via well-formed transactions. Subjects never touch objects directly; they pass through the access triple. It enforces separation of duties and auditing over Constrained Data Items. Brewer-Nash (Chinese Wall) dynamically blocks conflicts of interest, isolating datasets a consultant has already touched.

The state-machine model says a system staying in a secure state across every transition is secure. The reference monitor is the abstract concept enforcing this; its real implementation is the security kernel, which must be tamperproof, always invoked, and small enough to verify.

Cryptography

Think of cryptography like a postal system with locks. A symmetric lock uses one key both to seal and open the box, so whoever holds it can do both. Symmetric cryptography (AES, ChaCha20) is fast and ideal for encrypting large volumes, but distributing that shared key safely is the hard part. Asymmetric cryptography (RSA, ECC) solves distribution using a public/private key pair, but it is far slower. Real systems combine both: asymmetric exchanges a session key, then symmetric encrypts the actual traffic. TLS works exactly this way.

Hashing produces a fixed-length, one-way fingerprint (SHA-256) for integrity, never confidentiality. A digital signature hashes the message, then encrypts that hash with the sender's private key. Anyone with the public key verifies it, giving integrity, authentication, and the only mechanism delivering non-repudiation. PKI binds public keys to identities. A Certificate Authority (CA) issues X.509 certificates; clients check revocation via CRL (a published blacklist) or OCSP (a live lookup, faster and current). The key management lifecycle spans generate, distribute, store, use, rotate, archive, revoke, and destroy. Strong generation, escrow, and clean destruction matter as much as the algorithm.

Post-quantum: NIST finalized FIPS 203 (ML-KEM, key encapsulation, from CRYSTALS-Kyber) and FIPS 204 (ML-DSA, signatures, from Dilithium) in August 2024. These lattice-based schemes resist quantum attacks that will break RSA and ECC. CNSA 2.0 targets full migration by 2030, so "harvest-now-decrypt-later" risk is exam-relevant today.

System & physical security

Think of a hardened bank vault that still has one weak air vent on the roof. The walls are perfect, but attackers study the overlooked physical and timing gaps. System and physical security in CISSP Domain 3 is exactly this hunt for the overlooked gap.

System-level vulnerabilities exploit timing and shared resources. A TOCTOU flaw is a race condition where an attacker swaps a file between the permission check and the actual use. The exam fix is simple: shrink the check-to-use window, lock the resource, and validate atomically. A covert channel leaks data through paths never meant for communication. Storage channels hide data in shared temp space; timing channels signal secrets through clock-measurable delays. Side-channel attacks (Spectre, Meltdown, power and cache timing) infer keys without ever reading memory directly. Countermeasures include constant-time code, noise injection, and cache partitioning.

Cloud and virtualization security adds VM escape, hypervisor compromise, and noisy-neighbour resource leakage. You enforce isolation, patch the hypervisor first, and treat the shared-responsibility model as exam gospel. A TPM anchors secure boot and disk encryption keys to one machine. An HSM handles bulk cryptographic operations; for cloud, demand FIPS 140-2/140-3 Level 3 validation.

Physical and environmental controls follow Uptime Institute tiers: Tier I (basic), Tier II (redundant components), Tier III (concurrently maintainable, N+1), Tier IV (fault tolerant, 2N). Pair UPS plus generators for power, and use clean-agent fire suppression near hardware.

Domain 3 in the AI era (2026)

Domain 3 has always been about building trust into systems by design. In 2025-26 that mandate expanded to cover two new frontiers: securing the AI models themselves and defending today's encryption against tomorrow's quantum computers. A CISSP architect must now bake both into reference architectures from day one, not bolt them on later.

On the AI side, NIST finalised AI 100-2e2025 (March 2025), the authoritative taxonomy of adversarial ML. It classes attacks as evasion, poisoning, privacy, and (for GenAI) misuse — and for the first time covers LLMs, RAG pipelines, and autonomous agents. The design-layer lesson: a poisoned training set or a tampered model file is a supply-chain compromise, so model provenance, signed weights, and data-integrity controls belong in the architecture phase.

To protect models at runtime, architects turn to confidential computing. The NVIDIA H100/H200 were the first GPUs with a hardware TEE (AES-256 memory encryption + remote attestation), keeping model weights and inference inputs encrypted even from the cloud provider — at 95-99% of native speed.

For crypto futures, NIST's PQC standards FIPS 203 (ML-KEM), 204 (ML-DSA), and 205 (SLH-DSA) counter the harvest-now-decrypt-later threat. Since August 2025, Chrome and Firefox enable the hybrid X25519MLKEM768 handshake by default; the NSA's CNSA 2.0 sets 2030 for national-security systems, with US federal migration by 2035. Crypto-agility is now a hard architectural requirement.

ML-KEM TEE/attestation model poisoning

Strengths tip: If you score high on Strategic/Big-Picture thinking, you'll excel here — PQC migration is fundamentally an inventory-and-roadmap problem (find every use of vulnerable crypto first), not a single code fix.

The AI-era angle, in four cards

What 2026 adds to this domain — flip to see why each matters.