Secure System Design, Evaluation Models, and Physical Security

The two rules and the wall between them

An organization classifies its files into two buckets. The high bucket holds the products that have not been announced yet, the low bucket holds the public website, and a handful of people must work in both. The security team writes two rules. Rule one comes from the confidentiality side: classified material must never flow down into the public bucket. Rule two comes from the integrity side: unvetted material must never flow up into the product bucket, because someone who can edit the public site should not be able to plant changes inside the unreleased roadmap. Then the team looks at the result and finds the catch. Under rule one, nobody may copy a classified file down to the public site. Under rule two, nobody may copy an unvetted file up into the classified bucket. A document that must move from the roadmap into the announcement cannot move at all, because every move is blocked in one direction or the other.

That pinch, two correct rules that strangle a legitimate workflow when enforced together, is the heart of this chapter’s first half. Security models are not abstract decoration. They are formal answers to the question “what is a secure state, and which transitions keep the system in one?” The Bell-LaPadula model says a secure state is one where classified data cannot flow to people who may not see it. The Biba model says a secure state is one where untrusted data cannot corrupt trusted data. Both are right, both are published and formal, and a system that tries to enforce both against the same label orderings discovers that its users can no longer move anything. Understanding why the models disagree, and which model the exam expects for which goal, is what separates candidates who memorize definitions from candidates who can reason about a scenario.

The second half of the chapter follows the same material out of the mathematics and into the physical world. Evaluation schemes like the Common Criteria and FIPS 140 are the industrial version of “trust but verify”: they grade how much assurance a product’s claims carry, and the exam expects you to know what a grade means and, more importantly, what it does not mean. Hardware roots of trust such as the TPM and the HSM move the protection from the policy layer into silicon, where tampering costs more. And physical security, the layer most people think of as fences and guards, turns out to be the same design discipline as everything before it: layered controls, fail-safe defaults, and a clear idea of what each control is supposed to deter, detect, delay, or answer.

The chapter ends with practice questions. The models and evaluation material reward exact vocabulary, and the physical material rewards judgment about layered defense, so the questions mix the two.

What a security model is

A security model is a formal description of a security policy: a precise statement of the states a system may be in and the transitions between states that are allowed. The formality is the point. A policy document that says “protect sensitive data” cannot be enforced by a mechanism, because the sentence is not precise enough to check. A model says “a subject with clearance level Ls may read an object with classification level Lo only if Ls dominates Lo”, and that sentence can be enforced, tested, and proven. The reference monitor from the chapter on design principles is the enforcement side of the same coin: the model defines the rule, the reference monitor applies it to every access, and the trusted computing base is the set of components the rule depends on.

Three terms recur in every model and in every exam question about them. A subject is an active entity, usually a user or a process acting for a user. An object is a passive entity, a file, a database row, a record, a memory segment. A state is the complete set of subjects, objects, and their access relationships at a moment in time. A model is secure if every state reachable from a secure initial state through allowed transitions is itself secure, which is why the models are sometimes called state machine models: the security property is an invariant that every transition must preserve.

The models also share a common tool for expressing levels: the lattice. A lattice is a partially ordered set in which every two elements have a least upper bound and a greatest lower bound. Classifications such as unclassified, secret, and top secret form a simple chain, but real systems add categories, compartments such as “finance” and “personnel”, so that a level is a pair of a classification and a set of categories, and “dominates” means “classification at least as high and categories a superset of”. Lattices appear on the exam mostly as the structure underneath Bell-LaPadula and Biba, so the exam fact is modest: the levels form a lattice, and “dominates” is the ordering the models use.

With that vocabulary in place, the models themselves are the substance. They split into three families by the property they protect: confidentiality models, integrity models, and conflict-of-interest models, plus a small family of administrative models that describe how rights themselves are created, transferred, and removed. The exam asks three kinds of questions about them: which model protects which property, which property permits which flow, and which scenario the model was invented for.

Bell-LaPadula: the confidentiality model

Bell-LaPadula, named for David Bell and Leonard LaPadula and formalized in 1973 for the US Department of Defense’s time-sharing systems, is the model the exam treats as the default answer whenever the protected property is confidentiality. It protects secrets from leaking to people who are not cleared for them, and it was written to solve a specific problem that still drives exam questions: the Trojan horse.

The problem is this. A cleared user at Secret level opens an email attachment. The attachment contains a small program that, while running with the user’s rights, reads every Secret file it can reach and copies them to a public share. Nothing in an ordinary access control system stops it, because the program runs as the user, and the user is allowed to read Secret files and write to the public share. The system has no rule that connects the two permissions. Bell-LaPadula supplies the rule. It defines a dominance ordering over labels and then states three properties.

The simple security property, sometimes called the ss property, forbids read-up: a subject at a lower level cannot read an object at a higher level. This is the property that stops an unclassified user from reading a Secret file directly.

The star property, written with an asterisk in the original notation and called the star property ever since, forbids write-down: a subject at a higher level cannot write to an object at a lower level. This is the property that stops the Trojan horse. The program running as the Secret user can read Secret files, but it cannot write them to the public share, because that write crosses the boundary downward. The star property is the model’s crown jewel and the exam’s favorite point of confusion, because it is counterintuitive: it restricts the powerful subject, not the weak one. The user at the highest level is the one who loses the ability to write downward, precisely because the highest-level user is the one a Trojan horse would most want to impersonate.

The discretionary security property covers the discretionary part of the policy: access decisions made by users, expressed through an access control matrix, are allowed only when the mandatory properties above do not object. Some treatments also mention the strong star property, a stricter variant in which a subject may read and write only at its own level, which eliminates both read-up and write-up and is used when even controlled upward flows are unacceptable.

The exam questions on Bell-LaPadula almost always present a scenario and ask which property governs it: a process denied a read is the simple security property, a process denied a downward write is the star property, a system that allows a low-level user to post to a higher-level queue is the star property permitting write-up, which is allowed. That last case deserves emphasis because it is the model’s known weakness, and the exam likes to probe it. Bell-LaPadula permits write-up: a low-level subject may write to a higher-level object, because doing so cannot leak the high object’s contents downward. But write-up is exactly how untrusted data pollutes trusted objects. A web form that writes customer input into the production database is writing up in Bell-LaPadula terms, and the model is happy to permit it. Confidentiality models are silent on integrity, and this silence is the standard critique of Bell-LaPadula: it prevents the Trojan horse from leaking secrets down, but it does nothing to stop low-quality or hostile data from flowing up, and it does not address covert channels or availability at all.

Biba: the integrity mirror

Kenneth Biba’s 1977 model, written in response to exactly that silence, is the mirror image of Bell-LaPadula with integrity as the protected property. Where Bell-LaPadula stops untrusted readers from seeing high data, Biba stops untrusted data from corrupting high-integrity objects. It uses the same lattice machinery and the same style of property, inverted.

The simple integrity property forbids read-down: a subject with a lower integrity level cannot read an object with a higher integrity level. The reasoning is that reading untrusted data contaminates the reader: a process that ingests a corrupted configuration file cannot be trusted afterward, so the rule prevents the ingestion. The star integrity property forbids write-up: a subject at a lower integrity level cannot write to an object at a higher integrity level, which is the rule that keeps production systems from being modified by low-integrity processes. Biba also adds an invocation property: a low-integrity subject may not invoke or call a high-integrity subject, closing the route where a corrupted process commands a trusted one to act on its behalf.

The practical picture is a quarantine. Untrusted inputs, vendor code, email attachments, and user-generated content live at low integrity. Production binaries, financial records, and configuration live at high integrity. Biba says the two populations do not mix: low data never flows into high objects, and low processes never drive high processes. The model is elegant, and its exam value is partly that it is so cleanly the opposite of Bell-LaPadula: Bell-LaPadula says no read up and no write down, Biba says no read down and no write up. The moment an exam scenario mentions “integrity” or “untrusted input corrupting a system”, the answer is Biba. The moment it mentions a Trojan horse leaking secrets, the answer is Bell-LaPadula. Mixing the two up is the most common mistake on this material, and the questions at the end of the chapter drill the distinction.

Clark-Wilson: integrity for the commercial world

David Clark and David Wilson’s 1987 model exists because Biba, for all its elegance, is hard to apply to a business. Biba treats data as a ladder of trust and forbids upward flows, but real commercial systems are full of legitimate upward flows: customers type data into banking systems, clerks enter orders, and suppliers upload catalogs. Clark-Wilson provides the integrity model that a bank can actually run, and the exam presents it as the commercial counterpart to Biba’s military abstraction.

The model’s machinery is a small set of definitions. A constrained data item (CDI) is an object whose integrity the system must protect, such as a customer balance or an inventory count. An unconstrained data item (UDI) is everything else, raw input that has not been validated. An integrity verification procedure (IVP) is a procedure that checks that a CDI is in a valid state, that a balance still sums correctly, for example. A transformation procedure (TP) is the only thing that may change a CDI: any change to a constrained item must pass through an authorized TP, which validates its input and applies business rules before modifying the item. Users never touch CDIs directly; they invoke TPs, and a certified relation associates each user with the set of TPs that user may invoke.

Two design ideas sit on top of this machinery, and both are exam names. Well-formed transactions mean that data manipulation happens only through the TPs, so no user can write an arbitrary value into a CDI and every change carries the validation the business rules require. Separation of duties means that the user who creates a transaction cannot be the user who approves it: the certification of a TP and the authorization of a user to run it are separate relations, and the model requires that no single user holds all the rights needed to complete a sensitive transaction. The two-person rule in finance, where a transfer is entered by one officer and released by another, is Clark-Wilson in production, and the exam expects you to say so.

The question the exam uses to separate Biba from Clark-Wilson is usually “which integrity model is appropriate for a commercial application that must allow validated input to enter high-integrity systems?” Biba forbids the flow; Clark-Wilson structures it: raw input enters as a UDI, a TP validates it, and the validated result becomes part of a CDI. Both models protect integrity, but Biba is a pure lattice of trust levels while Clark-Wilson is a process discipline, which is why Clark-Wilson is the model that maps onto real banking, ERP, and payment workflows.

Brewer-Nash: the conflict-of-interest model

The Brewer-Nash model, published by David Brewer and Michael Nash in 1989 and usually called the Chinese Wall model, protects a different property: freedom from conflict of interest. It was written for the problem that consultancy and financial-services firms have, where one firm serves many clients who compete with one another. A consultant who works for Bank A and Bank B, or for two rival airlines, holds a position that invites abuse, deliberate or accidental: pricing data learned from one client inevitably shapes advice given to the other.

The model’s structure is simple. Objects are grouped into conflict classes, sets of data that must not be accessible together. The rule is that a subject may access an object only if the subject has never accessed another object in the same conflict class, and access decisions are based on the subject’s history of accesses rather than on a static label. The first time the consultant opens a file from the Acme Airlines dataset, they join that conflict class. From that moment, access to the Delta Airlines dataset is denied, even though the consultant’s clearance, their need, and the file’s classification would all allow it. The wall is built from history, and it only grows: the set of datasets a subject may touch shrinks with every dataset they touch.

The exam angle on Brewer-Nash is almost always recognition: a scenario about a consulting firm, an investment bank, or any multi-client environment where the forbidden thing is simultaneous access to two competitors’ data is a Chinese Wall question. The model also has one subtlety worth knowing: sanitized information, data that has been scrubbed of anything identifying the client, may be moved across the wall, because the conflict of interest lives in the association between data and client, not in the data itself.

The administrative models and the rest of the family

A second small family of models asks a different question: not which flows are allowed, but how rights are created, granted, transferred, and revoked. The exam touches these models lightly, usually with one definition question, but the definitions are stable and easy to bank.

The Graham-Denning model, from 1972, describes access control in terms of subjects, objects, and a set of primitive operations: create an object, create a subject, delete an object, delete a subject, grant a right, remove a right, transfer a right, and read a right. Its contribution is showing that rights themselves are managed by rules, that granting and transferring are operations that a model must regulate, and that the creator of an object owns the initial rights to it. The Harrison-Ruzzo-Ullman model, from 1976, builds on the access control matrix and commands that modify it, and its famous result is negative: the general safety problem, deciding whether any subject can ever acquire a right it does not hold, is undecidable. The exam sometimes asks that result directly, “which model proved that the safety problem is undecidable”, and the answer is HRU. The Take-Grant model treats rights as edges in a graph and defines take and grant rules for how edges move, a graph-theoretic approach whose main exam appearance is as a name in a list.

Two further concepts round out the family. Information flow models treat security as a property of where data travels rather than who accesses it, and Bell-LaPadula and Biba can both be read as flow models: they constrain the directions in which information may move. Non-interference, from Goguen and Meseguer, states the strongest version: actions by high-level subjects must not be observable by low-level subjects at all, a property far stronger than Bell-LaPadula, which the exam mentions mainly as a stricter ideal.

Combining models, and the manager’s view

Real systems do not run one model; they run combinations, and the exam asks what happens when combinations collide. The opening story is the collision: Bell-LaPadula forbids read-up and write-down, Biba forbids read-down and write-up, and applied to the same label ordering the two models leave a subject able to read and write only objects at its own level, an arrangement that protects everything and permits almost nothing. The resolution in practice is to use separate label spaces: classification labels for confidentiality, integrity labels for integrity, and objects carrying both. A system can then allow a Secret-cleared analyst to read Secret documents (allowed by Bell-LaPadula, and if the documents are high integrity, allowed by Biba too) while a low-integrity process still cannot write into the production database, because Biba is enforced on the integrity labels independent of the classification lattice.

The exam does not expect you to design such a system. It expects you to match models to goals: confidentiality to Bell-LaPadula, integrity to Biba or Clark-Wilson depending on whether the scenario is military or commercial, conflict of interest to Brewer-Nash, rights management to Graham-Denning and HRU. And it expects the manager-perspective judgment: a model is a policy formalized, it is enforced by a reference monitor, and no model, however elegant, supplies the human discipline around it. When a scenario asks “which model”, the discriminator is the property at risk, and when a scenario asks “what is missing”, the answer is often a second model covering the property the first one ignores.

Evaluation: what assurance means, and who certifies it

The models say what a system should do. The evaluation industry answers a different question: how much confidence can you have that a product actually does it? The exam vocabulary splits into functionality and assurance. Functionality is what the product does, the features and controls it provides. Assurance is the evidence that the functionality is correctly implemented and protected from tampering. A product can have rich functionality and thin assurance, a feature list is not proof, and the entire evaluation apparatus exists to make assurance measurable.

The historical line begins with the US Trusted Computer System Evaluation Criteria, TCSEC, published in 1985 as DoD 5200.28-STD and known from the color of its cover as the Orange Book. TCSEC defined classes from D, minimal protection, through C1 and C2, the discretionary classes, through B1, B2, and B3, which introduced mandatory access control with labels and increasingly rigorous design review, to A1, verified design, which required a formal model and its proof. C2 is the class the exam is most likely to name, because it demanded individual accountability, audit, and object reuse protection. Europe answered with ITSEC, the Information Technology Security Evaluation Criteria, whose levels ran E0 through E6. Both lines, plus work in Canada, were merged into the Common Criteria, and that is the standard the current exam expects you to know cold.

The Common Criteria is published as ISO/IEC 15408, currently at version 3.1 revision 5. It is not a list of required controls; it is a language for expressing security requirements and a framework for evaluating whether a product meets the requirements it claims. The vocabulary is small and exact. The target of evaluation (TOE) is the product under evaluation. A protection profile (PP) is an implementation-independent statement of security requirements for a category of products, “a firewall for a small office” or “a database for a government environment”, written so that any product claiming conformance must meet the same bar. A security target (ST) is the TOE-specific document that states what this particular product does, which requirements it claims, and how, and it is the document the evaluation actually checks. Security functional requirements (SFRs) are the functional claims and security assurance requirements (SARs) are the evidence requirements, and both are drawn from the standard’s catalog of components.

The evaluation itself produces an assurance level, EAL1 through EAL7. The names are the exam fact: EAL1, functionally tested; EAL2, structurally tested; EAL3, methodically tested and checked; EAL4, methodically designed, tested, and reviewed; EAL5, semiformally designed and tested; EAL6, semiformally verified design and tested; EAL7, formally verified design and tested. Two exam traps sit on top of these names. First, a higher EAL is more assurance, not more security functionality: an EAL7 product can do less than an EAL4 product, it is simply proven more rigorously. Second, an evaluation certifies the product against its own security target, not against your environment: the evaluation says the product meets the claims it made, in the configuration evaluated, and no more.

The international machinery matters only insofar as the exam mentions it. Evaluations are run by licensed laboratories and overseen by national schemes, in the United States the National Information Assurance Partnership, NIAP. Mutual recognition among countries runs through the Common Criteria Recognition Arrangement, CCRA, which since its 2012 revision recognizes evaluations only at EAL2 through EAL4 and under collaborative protection profiles, rather than at the higher levels. The practical reading: a product evaluated to EAL4 by one scheme is accepted by the other CCRA members, and a product evaluated to EAL7 is not part of that mutual recognition.

FIPS 140: the cryptographic module standard

The Common Criteria evaluates whole products. FIPS 140 evaluates the smaller thing inside them: the cryptographic module. The distinction is a favorite exam discriminator. A TLS appliance might carry both a Common Criteria evaluation of the appliance and a FIPS 140 validation of its crypto module, and the two certifications answer different questions about different objects.

FIPS 140-2, published in 2001, defined four security levels for cryptographic modules. FIPS 140-3, published in March 2019 and based on ISO/IEC 19790, carries the same four levels forward, and the exam expects the level names and the boundary between them. Level 1 is the baseline: production-grade components and a defined boundary, but no special physical security. Level 2 adds tamper-evidence, tamper-evident seals or coatings that show when the module was opened, and role-based authentication, an operator authenticating as holding a role rather than as a named individual. Level 3 adds tamper-resistance, mechanisms that block or severely impede physical access, identity-based authentication that ties actions to a specific named operator, and stronger protection of the module’s interfaces. Level 4 is the top: full tamper protection including detection and response, with the module zeroizing keys when attack is detected, plus environmental failure protection that responds to voltage and temperature extremes.

The validation machinery has its own name. The Cryptographic Module Validation Program, CMVP, run jointly by NIST and the Canadian Centre for Cyber Security, tests modules and issues certificates, and a “FIPS validated” claim means a module holds a CMVP certificate for a specific firmware version. The word version is the trap: validation attaches to an exact build and configuration, so the same hardware with different firmware is a different module needing its own validation. A second useful date: FIPS 140-2 certificates were allowed to run out, and the last of them expire on September 22, 2026, which means that in current designs the standard to name is FIPS 140-3. Payment HSMs and the key-management hardware in this chapter’s next section are the products you will most often see cited against it.

Two other FIPS publications complete the exam’s picture. FIPS 199 defines security categorization, mapping the impact of loss of confidentiality, integrity, or availability to low, moderate, or high, the scheme that drives SP 800-53 control baselines. FIPS 201, the Personal Identity Verification standard, defines the PIV card used for federal identity, and it matters to this chapter because PIV is a physical-access artifact as much as a logical one: the badge at the door and the credential at the workstation are the same token.

Hardware security: roots of trust in silicon

Software can be patched. Hardware cannot, not in the field, and that asymmetry is why the strongest guarantees in a system live in hardware. The modern architecture of trust starts with a hardware root of trust, a tamper-resistant component whose trustworthiness is taken as given and from which everything else derives. The exam knows three of these components well: the TPM, the HSM, and the secure enclave, plus a supporting cast of physical unclonable functions and tamper mechanisms.

The Trusted Platform Module, TPM, is a dedicated crypto coprocessor standardized by the Trusted Computing Group and published as ISO/IEC 11889, with TPM 2.0 the current version and a requirement for Windows 11 certification. Its job is to be the platform’s identity and honesty meter. The TPM holds a small number of shielded storage slots, keys that never leave the chip, and a set of platform configuration registers, PCRs, that record hashes of the boot chain. Measured boot, or secure boot in its UEFI variant, is the process that fills them: at each boot stage, the stage measures the next stage by hashing it, and extends the result into a PCR. An attacker who swaps the bootloader changes the hash, changes the PCR value, and leaves a fingerprint the TPM can attest to.

Three TPM capabilities carry exam weight. Sealed storage binds a secret to platform state: the TPM releases a key, a disk-encryption key, for example, only if the PCR values match the values recorded when the secret was sealed, so a machine whose boot chain has been modified cannot unlock its own disk. Remote attestation is the reporting side: the TPM signs the current PCR values with a key, an attestation identity key derived from the endorsement key burned into the chip at manufacture, and a remote verifier, a VPN gateway or an inventory system, can cryptographically check that the machine booted the approved software before granting it network trust. The endorsement key and the identity keys give the platform a manufactured identity that cloning would break, which is why a genuine TPM is hard to fake and why TPM questions on the exam usually resolve to “the chip that measures the boot chain, seals secrets to it, and attests it to others”.

The Hardware Security Module, HSM, is the TPM’s big sibling, a dedicated appliance for the most sensitive cryptographic duties: certificate authority private keys, payment-card keys, root keys of trust hierarchies, and key ceremonies. Where the TPM is a cheap chip inside every laptop, the HSM is a purpose-built box, typically validated at FIPS 140-3 Level 3 or higher, that performs crypto operations without ever exporting the keys in the clear. The exam vocabulary around HSMs is key management discipline. Split knowledge means the secret is divided so that no single person holds all of it, and dual control means two or more authorized people must act together before a sensitive operation, key generation, key backup, key destruction, runs. The ceremony around a root CA key, where several officers each enter a portion of a passphrase and the HSM requires two of them to act, is the physical form of separation of duties, and the chapter on cryptography touched the same ideas from the key-management side.

Between the TPM and the HSM sit the secure enclaves and secure elements: hardware-isolated regions inside a processor, such as Intel SGX, AMD SEV, and Arm TrustZone, that protect code and data from the rest of the operating system, and the tamper-resistant secure elements inside phones and payment devices that hold the keys for mobile wallets. These are the hardware version of the security kernel from the design-principles chapter: a small, trusted, isolated place that the rest of the system cannot reach. Physical unclonable functions, PUFs, round out the family by deriving a unique device identity from microscopic manufacturing variations in silicon, so that each chip effectively has a fingerprint that cannot be cloned, used for anti-counterfeiting and for binding keys to a specific device.

The hardware layer also has an attack surface, and the exam expects you to know the attacks by name. Cold boot attacks exploit DRAM remanence, the property that memory contents persist for a time after power loss, so an attacker who freezes a running machine, yanks the power, and reads the memory with another machine can recover encryption keys that were resident in RAM; researchers at Princeton demonstrated the technique in 2008. DMA attacks exploit direct memory access, the bus feature that lets peripherals such as FireWire and Thunderbolt devices read memory without CPU mediation, which is why those ports are disabled on sensitive machines and why input/output memory management units, IOMMUs, exist. Debug interfaces such as JTAG, left enabled in production, give an attacker a direct hardware back door, and fault injection, glitching voltage or clock lines to make a chip skip a check, is how some of the most celebrated hardware attacks work. The supply chain is the deepest version of the same problem: a chip modified at the factory cannot be healed by a software patch, which is why firmware signing, attested boot, and source assurance for hardware matter to the manager. The decision logic the exam rewards is simple: software keystores are weakest, a TPM protects a platform identity and its boot state, and an HSM is where the keys that must survive and be strongly protected live.

Physical security: deter, detect, delay, respond

The last layer in the chapter is the one every security career starts with and every exam underestimates. Physical security is defense in depth applied to space and time. The organizing frame is a short list of verbs: deter, detect, delay, respond. A fence and lighting deter, they persuade a casual intruder to pick an easier target. Sensors, cameras, and guards detect, they notice that an intrusion is happening. Locks, mantraps, and barriers delay, they buy the minutes that response needs. Guards and response teams answer, they turn detection into intervention. A physical security program is judged by whether each layer has a job and whether the layers reinforce one another, and the exam questions that ask “which control is missing” or “which control belongs to which layer” are testing exactly this frame.

The standards anchoring the frame are the ones the exam likes to see named. NIST SP 800-53’s Physical and Environmental Protection family carries the controls: PE-2 for physical access authorizations, PE-3 for physical access control, PE-6 for monitoring physical access, PE-8 for visitor access records, PE-9 for power equipment and cabling, PE-10 for emergency shutoff, PE-11 for emergency power, PE-12 for emergency lighting, PE-13 for fire protection, PE-14 for temperature and humidity, PE-15 for water damage protection, PE-16 for delivery and removal, and PE-18 for location of system components. ISO/IEC 27001:2022 Annex A compresses the same discipline into controls 7.1 through 7.14, covering the physical security perimeter, physical entry, secure offices and rooms, physical security monitoring, protection against physical and environmental threats, working in secure areas, clear desk and clear screen, equipment siting, off-premises assets, storage media, supporting utilities, cabling security, equipment maintenance, and secure disposal. A candidate who can map a physical control in a scenario to the PE control or the Annex A control that names it has answered the way the exam rewards.

The design starts before the building: site selection and CPTED. Crime Prevention Through Environmental Design holds that the physical environment itself shapes behavior. Its four strategies are natural surveillance, arranging windows, lighting, and sight lines so that intruders are visible and hiding places are few; natural access control, guiding people through visible, single-entry routes; territorial reinforcement, using fences, landscaping, and signage to signal that an area is owned and watched; and maintenance, keeping the environment visibly cared for, because neglect signals that nobody watches. Site selection extends the same logic: flood plains, crime statistics, proximity to hazards, and the threat model of the facility are decisions made before the fence is drawn.

The perimeter is the first line. Fences mark the boundary and deter, bollards stop vehicles, gates control vehicle entry, and exterior lighting removes darkness as cover. Entry is where access actually happens. Mantraps, interlocking two-door vestibules where only one door opens at a time, hold a person in a chamber while their credentials are checked and while guards decide, and they are the classic exam answer for “how do you stop tailgating through a secure door”. Badge readers, proximity cards, and biometric readers authenticate entry, and in federal contexts the PIV card of FIPS 201 is the badge itself, one credential for door and workstation. The interior continues the layering: locks on every door, cages and safes for sensitive media, alarmed server rooms, and clear rules for visitors, escorts, and the two-person rule in the most sensitive areas. Guards are the most flexible control in the inventory, able to judge, question, and respond where machines cannot, and their patrols and response protocols are themselves a documented control. Cameras and motion sensors, the detection layer, need their own policy: what is covered, how long footage is retained, who may view it, and how it is protected as personal data under privacy law.

Two terminology traps deserve their own paragraphs because the exam sets them repeatedly. The first is fail-safe versus fail-secure on doors and locks. In physical access control, fail-secure means the door stays locked when power fails, and fail-safe means the door opens when power fails. The names come from the two priorities: fail-secure protects the room, fail-safe protects the people, and life safety wins, so emergency exits and the doors of occupied spaces fail open so that nobody is trapped in a fire or a blackout. The design-principles chapter used fail-safe defaults to mean “deny by default”, and the exam expects you to keep the contexts apart: in policy, fail-safe means fail closed; in physical life safety, fail-safe means fail open. The second trap is power. A blackout is total loss of power, a brownout is a prolonged reduction in voltage, a sag is a brief one, a spike is a brief surge above normal, and noise is electromagnetic interference on the line. The layering of controls is the exam fact: a UPS, ideally the online double-conversion kind, conditions power and carries the load through short interruptions; a generator, with fuel and a transfer switch, covers extended outages; and the emergency power off, EPO, the big red button, kills everything at once, which is why EPO switches exist for firefighter safety and why their placement, training, and accidental-use consequences are a control design problem themselves. Redundancy is expressed in the arithmetic of capacity: N+1 means one more unit than the load needs, and 2N means twice as much, the difference between surviving a component failure and surviving a whole redundant path.

The environmental half of physical security protects the equipment from the building and the building’s own systems. Fire protection starts with classification: class A fires burn ordinary combustibles like wood and paper, class B burns flammable liquids, class C is energized electrical equipment, class D is combustible metals, and class K is cooking oils. The exam fact is which suppressant fits which class: water on a class A fire, carbon dioxide or dry chemical on class C, and never water on an energized electrical fire. In a data center the answer is a clean-agent system: gases such as FM-200, Novec 1230, and Inergen that suppress fire without leaving residue and without conducting electricity, replacing the halon that the Montreal Protocol banned from new production because it destroyed the ozone layer. Sprinkler types also have names: wet pipe keeps water in the pipes, dry pipe holds air until a head opens, and pre-action requires two detection signals before water is released, which is why pre-action is the acceptable water option for computer rooms, it reduces accidental discharge to nearly zero. Detection matters as much as suppression: smoke detectors, heat detectors, flame detectors, and aspirating systems that sample the air and catch a fire before visible smoke, and water-leak sensors on raised floors that catch the broken pipe that suppression systems exist to fight. Temperature and humidity round out the environment: HVAC keeps the room in the range the equipment is designed for, roughly 18 to 27 degrees Celsius with humidity high enough to avoid static discharge, and the monitoring of both is a control, PE-14 in the NIST family, not a comfort amenity.

The last idea in physical security is convergence. The badge reader at the door and the identity provider in the network are the same system; door events should reach the same SIEM as login failures, camera retention should follow the same data-retention policy as logs, and an access-card revocation should be nearly immediate, because the badge the departing employee still holds is a logical credential that happens to live on plastic. The exam rewards candidates who notice that physical and logical controls enforce the same policy through different mechanisms, and the strongest answers treat the building as one more layer of the architecture rather than as a separate discipline.

How the exam thinks about this material

Three patterns cover most of the questions this chapter feeds. The first is property matching. The stem names a property, confidentiality, integrity, conflict of interest, and the right answer names the model that protects it: Bell-LaPadula for confidentiality, Biba or Clark-Wilson for integrity with the commercial/military discriminator, Brewer-Nash for conflict of interest, HRU for the safety problem. The second is the assurance vocabulary. A protection profile is requirement-for-a-category, a security target is claims-for-this-product, an EAL is assurance and not functionality, and FIPS 140 is for the crypto module while Common Criteria is for the whole product. The third is layered physical judgment. Physical questions are decided by the layer a control belongs to, life safety prevailing over room security, clean agents over water in a server room, UPS before generator in the power chain, and every single answer that relies on one control doing everything is wrong for the same reason every single-principle answer is wrong elsewhere in this book: defense in depth is the house style, and the exam is written in that style.

Practice questions

1. A cleared analyst at Secret level opens a command window and tries to read a Top Secret file. The file system denies the read. Which model property is enforcing the denial?

A. The star property of Bell-LaPadula B. The simple security property of Bell-LaPadula C. The star integrity property of Biba D. The discretionary security property of Bell-LaPadula

2. A malicious spreadsheet at Secret level contains a macro that opens every Secret file it can reach and copies them into a public shared folder. The copies fail. Which property blocks them?

A. The simple security property, because the macro may not read the Secret files B. The star property, because a Secret subject may not write down to the public folder C. The simple integrity property, because the macro may not read untrusted data D. The invocation property, because the macro may not invoke a public subject

3. A system enforces Bell-LaPadula and nothing else. Which failure is the model’s known blind spot?

A. A high-level process reading low-level data and learning public facts B. A low-level process writing unvalidated data up into a high-level object C. A high-level process writing classified data down into a low-level object D. A low-level process reading a high-level object through a permitted share

4. A production analytics job must not read configuration files that have not been vetted, because a corrupted config could poison its output. Which Biba property states the rule?

A. The simple integrity property, no read down B. The star integrity property, no write up C. The invocation property, no invoking higher subjects D. The simple security property, no read up

5. A low-integrity web form attempts to write directly into the high-integrity production database. Under Biba, which property denies the write?

A. The simple integrity property B. The star integrity property C. The discretionary security property D. The strong star property

6. In the Clark-Wilson model, which component is responsible for checking that a constrained data item is still in a valid state?

A. A transformation procedure B. An integrity verification procedure C. A constrained data item’s own audit log D. The reference monitor’s access matrix

7. A payment system requires that a transfer be entered by one clerk and approved by a different clerk, and neither clerk may perform both steps. Which pair of Clark-Wilson concepts does this implement?

A. Constrained and unconstrained data items B. Well-formed transactions and separation of duties C. Lattice ordering and the simple integrity property D. Certified relations and integrity verification procedures

8. A consultancy that advises one bank may not access the records of a second bank, because both banks belong to the same conflict class. Which model states this rule?

A. Bell-LaPadula B. Biba C. Brewer-Nash, the Chinese Wall model D. Clark-Wilson

9. Under the Chinese Wall model, an analyst who has already opened a file from the Acme dataset is later denied access to a file from the Delta dataset, even though clearance, classification, and need all permit it. What makes the denial correct?

A. Delta is classified higher than the analyst’s clearance B. The analyst’s history of accesses places Acme and Delta in conflict C. The star property forbids writing down between the two datasets D. The integrity levels of the two datasets are incompatible

10. Which model is famous for proving that the general safety problem, deciding whether a subject can ever acquire a right it does not hold, is undecidable?

A. Graham-Denning B. Take-Grant C. Harrison-Ruzzo-Ullman D. Brewer-Nash

11. A model describes access control as subjects, objects, and a primitive set of operations that create and delete objects and subjects and that grant, remove, transfer, and read rights. Which model is this?

A. Graham-Denning B. Non-interference C. Biba D. The Chinese Wall model

12. An architect applies Bell-LaPadula and Biba to the same ordering of labels, so that the classification levels and the integrity levels are the same chain. What is the practical consequence for a subject at level X?

A. The subject may read and write objects at every level, because the two models cancel B. The subject may read and write only objects at level X C. The subject may read anything but write nothing D. The subject may write anything but read nothing

13. A government agency publishes a document that states the security requirements for any firewall it will purchase, with no product named. What is this document under the Common Criteria?

A. A security target B. A protection profile C. An evaluation assurance level D. A security functional requirement

14. In a Common Criteria evaluation, which document states the security requirements and claims of the specific product being evaluated, and is the basis of the evaluation?

A. The protection profile B. The security target C. The certificate of conformance D. The EAL specification

15. A vendor markets a product evaluated to EAL4 and compares it to a competitor’s EAL2 product by claiming twice the security. Why is the claim wrong?

A. EAL4 products are required to contain more security functionality than EAL2 products B. A higher EAL certifies more assurance, not more functionality, and the products’ claims were checked against different targets C. EAL levels apply only to cryptographic modules, not to whole products D. EAL4 is the maximum level, so no product can exceed it

16. A cryptographic module carries tamper-evident seals and requires an operator to authenticate as holding a specific role before use. Which FIPS 140-3 security level does this describe?

A. Level 1 B. Level 2 C. Level 3 D. Level 4

17. Which statement correctly distinguishes the Common Criteria from FIPS 140?

A. Common Criteria validates cryptographic modules; FIPS 140 evaluates whole products B. FIPS 140 validates cryptographic modules; Common Criteria evaluates whole products against their own security targets C. Both evaluate whole products, but FIPS 140 is limited to the United States D. Both validate cryptographic modules, but Common Criteria is international

18. A laptop’s TPM holds hashes of the UEFI firmware, the bootloader, and the kernel, and the full-disk encryption key is released only when those hashes match the values recorded at enrollment. What is this mechanism called?

A. Remote attestation with an endorsement key B. Sealed storage bound to platform configuration registers C. A key ceremony with dual control D. A hardware security module in Level 3 mode

19. A security team wants to verify that an endpoint is running the approved operating system before it joins the VPN. Which TPM capability directly supports this check?

A. Sealed storage B. Remote attestation, in which the TPM signs its current platform measurements C. The endorsement key alone, which proves the chip is genuine D. Measured boot without any reporting mechanism

20. A certificate authority stores its private key on a hardware security module. To generate the key, two officers must each supply part of a passphrase, and no single officer can reconstruct it. What are these requirements called?

A. Dual control and split knowledge B. Role-based authentication and tamper evidence C. Sealed storage and attestation D. Defense in depth and least privilege

21. An attacker gains a few minutes of physical access to a running server, cools the RAM, removes power, and reads the memory chips with another machine, recovering encryption keys that were resident in memory. Which control most directly reduces the risk from this attack?

A. A longer disk-encryption passphrase B. Restricting boot to approved devices and clearing memory on reboot C. Storing the key in the TPM instead of memory D. Using AES-256 instead of AES-128

22. A fire ignites in energized electrical equipment in a data center. Which suppression approach is the best primary choice for this environment?

A. Wet-pipe sprinklers B. A clean-agent gas system C. Class D dry powder D. Halon from the existing legacy cylinders

23. Line voltage drops below the nominal level for several minutes during a heat wave, and sensitive equipment starts failing. What is this power condition called, and what should carry the load through it?

A. A spike, handled by surge suppressors B. A brownout, handled by the UPS until the generator takes over C. A blackout, handled by emergency lighting D. Noise, handled by power conditioning alone

24. A server-room door is configured so that it remains locked when building power fails, because the room contains equipment whose compromise is worse than being locked out of it. What is this configuration called, and what is the life-safety rule that governs such decisions?

A. Fail-safe, and occupied spaces must fail open to protect people B. Fail-secure, and emergency exits and occupied spaces must fail open so nobody is trapped C. Fail-open, and the rule is that security always beats convenience D. Fail-closed, and the rule is that doors must never open automatically

25. A facility manager installs bollards at the loading dock, badge readers on every door, motion cameras in the corridors, and a guard team with a documented response protocol. An intruder climbs the fence at 2 a.m. and crosses the yard to a side door. Which control in the arrangement is primarily a delay control?

A. The motion cameras B. The badge readers, locks, and mantrap at the doors C. The guard team D. The bollards at the loading dock

Answers and rationales

1. B. The simple security property of Bell-LaPadula states that a subject at a lower level may not read an object at a higher level, which is exactly the denied read in the scenario. The star property governs writes (option A), Biba’s star integrity property governs untrusted writes upward (option C), and the discretionary property governs discretionary grants, not the mandatory classification denial (option D).

2. B. The macro runs with the analyst’s Secret rights and can read the Secret files, so the simple security property is satisfied (option A is wrong on the facts). The denial comes from the star property, which forbids a Secret subject from writing down to the public folder. Options C and D belong to the integrity family, which is not what the macro violates.

3. B. Bell-LaPadula permits write-up: a low-level subject may write to a higher-level object, because doing so cannot leak the high object’s contents downward. That permission is the known integrity blind spot, since unvalidated or hostile data can flow up into trusted objects. A high-level process reading low data is permitted and harmless (option A), the write-down in option C is exactly what the star property blocks, and a direct read-up in option D is blocked by the simple security property.

4. A. The simple integrity property of Biba forbids a subject from reading objects at a lower integrity level, on the theory that reading untrusted data contaminates the reader. The production job refusing to ingest unvetted configuration is that rule in action. The star integrity property governs writes (option B), the invocation property governs calling higher subjects (option C), and no read up is a Bell-LaPadula rule (option D).

5. B. The star integrity property of Biba states that a subject at a lower integrity level may not write to an object at a higher integrity level, so the low-integrity form’s write to the production database is denied. The simple integrity property governs reads (option A), the discretionary security property is Bell-LaPadula’s (option C), and the strong star property is a Bell-LaPadula variant (option D).

6. B. An integrity verification procedure is the Clark-Wilson component that checks whether a constrained data item is in a valid state. Transformation procedures change CDIs, but only through certified, validating logic (option A), CDIs do not audit themselves (option C), and the reference monitor enforces policy rather than validating business data states (option D).

7. B. Requiring different people to enter and approve a transaction is separation of duties, and requiring all data manipulation to pass through the approved transformation procedures is well-formed transactions. The other pairs name real Clark-Wilson machinery, but not the concept the scenario demonstrates (options A, C, and D).

8. C. Brewer-Nash, the Chinese Wall model, groups objects into conflict classes and forbids a subject from accessing objects in two conflicting sets, which is the rule that protects a multi-client consultancy from conflicts of interest. Bell-LaPadula and Biba operate on classification and integrity labels (options A and B), and Clark-Wilson protects commercial integrity through procedures (option D).

9. B. The Chinese Wall model is history-based: once the analyst accesses a dataset in a conflict class, access to the other members of that class is denied, regardless of clearance, classification, or need. The denial does not depend on the labels in options A, C, or D.

10. C. Harrison, Ruzzo, and Ullman formalized access control matrices with commands and proved that the general safety problem is undecidable. Graham-Denning is the earlier primitive-rights model (option A), Take-Grant models rights as graph edges (option B), and Brewer-Nash protects conflict of interest (option D).

11. A. The Graham-Denning model describes subjects, objects, and a primitive set of operations for creating and deleting them and for granting, removing, transferring, and reading rights. The other options are not administrative models of this kind (options B, C, and D).

12. B. Applied to the same label chain, Bell-LaPadula forbids reading up and writing down, while Biba forbids reading down and writing up, so the only flows that survive both models are reads and writes at the subject’s own level. The models do not cancel (option A), and neither direction becomes fully open (options C and D). Real designs separate the label spaces so each model governs its own ordering.

13. B. A protection profile is an implementation-independent statement of security requirements for a category of products, written before any product is chosen. A security target states the claims of one specific product (option A), an EAL is the resulting assurance grade (option C), and a security functional requirement is a component of the claims, not the whole document (option D).

14. B. The security target states the security requirements and claims for the specific target of evaluation and is the document the evaluation checks. The protection profile is the category-level requirements document (option A), and the other options are not documents of this kind (options C and D).

15. B. A higher EAL certifies more assurance, more rigor in design, testing, and review, not more functionality, and each evaluation is against that product’s own security target, so cross-product comparisons of “security” from EAL numbers are invalid. The other statements misstate what an EAL is (options A, C, and D).

16. B. FIPS 140-3 Level 2 adds tamper-evidence and role-based authentication. Level 1 has no such requirements (option A), Level 3 requires tamper-resistance and identity-based authentication (option C), and Level 4 adds tamper-protective wrapping and environmental failure protection (option D).

17. B. FIPS 140 validates cryptographic modules against a fixed level scheme, and the Common Criteria evaluates whole products against the requirements in their own security targets. The other pairings are reversed or incorrect (options A, C, and D).

18. B. Sealing binds a secret to platform state: the TPM releases the key only when the platform configuration registers match the values recorded when the key was sealed. Attestation is the reporting side (option A), key ceremonies involve people (option C), and the laptop’s TPM is not an HSM appliance (option D).

19. B. Remote attestation has the TPM sign its current platform configuration register values with a key derived from its endorsement key, so the verifier can cryptographically confirm which software booted. Sealed storage releases secrets (option A), the endorsement key proves chip authenticity but not boot state (option C), and measured boot without reporting gives no verifiable evidence (option D).

20. A. Split knowledge divides a secret so no single person holds it all, and dual control requires two or more authorized people to act together, the discipline used in HSM key ceremonies. The other options name unrelated mechanisms (options B, C, and D).

21. B. Cold boot attacks recover keys from memory remanence, so the most direct mitigation is preventing the attacker from gaining a bootable memory image: locking the boot order to approved devices and clearing memory on reboot. A longer passphrase does not help because the attack takes the key from RAM, not from guessing (option A), the key must be in memory while the system runs regardless of where it is stored at rest (option C), and AES-256 versus AES-128 does not change the memory exposure (option D).

22. B. For a data center fire, a clean-agent gas system suppresses without water, without residue, and without conducting electricity, which is why FM-200, Novec 1230, and Inergen are the standard answer. Pre-action sprinklers reduce accidental discharge but still use water (option A), class D powder is for combustible metals (option C), and halon production was banned under the Montreal Protocol, so legacy cylinders are a phase-out problem, not a design choice (option D).

23. B. A prolonged reduction in line voltage is a brownout, and the layered answer is that the UPS carries the load and conditions power while the generator starts and takes over for the extended outage. A spike is a brief overvoltage (option A), a blackout is a total loss (option C), and power conditioning alone does not provide ride-through (option D).

24. B. The configuration is fail-secure, the door stays locked when power fails, protecting the room. The governing rule is life safety: emergency exits and occupied spaces must fail open, fail-safe, so people are never trapped, which is why the scenario’s logic fits a room that is unoccupied when it matters and why the design must be justified against that rule. Options A, C, and D misstate both terms and the rule.

25. B. Locks, badge readers, and the mantrap are delay controls: they buy the time response needs. Cameras are detection controls (option A), the guard team is the response layer (option C), and bollards at the loading dock are vehicle barriers that belong to the perimeter deterrence layer, not to the path the intruder actually took (option D).

Security models, evaluation, and physical security on one page

Hold three threads and the chapter hangs together. Thread one is the property each model protects. Bell-LaPadula protects confidentiality with the simple security property, no read up, and the star property, no write down, and its blind spot is integrity, since write-up is permitted. Biba protects integrity with the simple integrity property, no read down, the star integrity property, no write up, and the invocation property, and it is the mirror of Bell-LaPadula applied to the same labels: they pinch a subject down to its own level. Clark-Wilson is the commercial integrity model, built from constrained and unconstrained data items, integrity verification procedures, transformation procedures, well-formed transactions, and separation of duties. Brewer-Nash, the Chinese Wall, protects against conflicts of interest using history-based access decisions over conflict classes. Graham-Denning formalizes the primitive rights, and Harrison-Ruzzo-Ullman proved the safety problem undecidable. Thread two is evaluation vocabulary. Assurance is evidence, not features. The Common Criteria, ISO/IEC 15408 version 3.1 revision 5, evaluates a target of evaluation against a security target, and protection profiles state category-level requirements; EAL1 through EAL7 grade assurance rigor, not functionality, and CCRA mutual recognition covers EAL2 through EAL4 and collaborative profiles. FIPS 140-3, based on ISO/IEC 19790, validates cryptographic modules at four levels, from Level 1’s production-grade baseline through Level 2’s tamper evidence and role-based authentication, Level 3’s tamper resistance and identity-based authentication, and Level 4’s tamper protection with zeroization and environmental failure protection, with the last FIPS 140-2 certificates expiring September 22, 2026. Thread three is physical security as layered space: deter with CPTED, site selection, fences, and lighting; detect with cameras, sensors, and guards; delay with locks, mantraps, and barriers; respond with guards and procedures, all mapped to NIST SP 800-53’s PE family and ISO/IEC 27001:2022 Annex A controls 7.1 through 7.14. Environmental controls finish the layer: clean-agent suppression in data centers, pre-action water systems where water is unavoidable, UPS and generator layering with EPO for emergencies, and the fail-safe versus fail-secure rule, where life safety wins and occupied spaces fail open. In hardware, the TPM measures the boot chain, seals secrets to the measurements, and attests them remotely, the HSM guards the keys that must never leave the box under split knowledge and dual control, and cold boot, DMA, debug-port, fault-injection, and supply-chain attacks are the threats the hardware layer exists to raise the cost of.

The next chapter, the Domain 3 practice test, will ask twenty-five questions across the whole domain, and this chapter is half of its material: the models, the evaluation schemes, and the physical layer sit beside the design principles and the cryptography that chapters twelve and thirteen taught. Before you go, hold the sentence this chapter keeps returning to: a model tells you what a secure state is, and the room, the chip, and the door are where the state becomes real.