Key Takeaways
Treat every network as hostile, including the one behind your firewall
Perimeters have a blind spot. Traditional networks divide systems into zones and inspect traffic mainly where it crosses between them. That approach grew from a time when private networks were relatively isolated. As organizations connected them to the internet, first through a few servers and later through broad outbound access, the boundary became porous. Malware can initiate an outward connection, receive commands, and use a compromised employee computer to move sideways through the network.
A breach can travel inward. The book traces a plausible path from a phishing email to an intern’s computer, a developer’s stolen credentials, and a production database. Firewalls may be correctly configured and still fail to stop this chain because hosts inside a zone are often trusted to communicate. Zero trust starts instead by treating internal traffic as exposed to hostile actors.
The argument shifts security from keeping intruders out to limiting what an intruder can do after entry. That aligns with resilience engineering, which plans for component failure rather than assuming failure can be eliminated. Network zones can still help contain broad classes of risk, but they are weak as identity proofs. A useful distinction is that a perimeter can reduce exposure while endpoint-level checks constrain authority. The book’s strongest point is not that boundaries are worthless, but that crossing one should never grant a lasting presumption of trust.
Authorize the whole request, not a username or IP address alone
Context changes access. A network agent is an on-demand collection of information about the user, application, and device involved in a request. It may include identity, role, device status, location, and a trust score. Authentication establishes who or what presented credentials; authorization decides whether that authenticated context may perform this particular action. Keeping the two separate prevents a valid login from becoming a blank cheque.
Assess each request. A company laptop might be permitted to reach source code while the same employee’s phone is not. Trust can also change with behavior: an unfamiliar location or an unusual request may trigger an added authentication step or denial. Agent data can be incomplete or self-reported, so policies should distinguish verified facts from weaker signals rather than treating every field as equally reliable.
The agent model resembles contextual access control: permission depends on the combination of actor, device, and action, not a single label. This is a meaningful defense against stolen credentials because an attacker may possess a password without controlling an enrolled device. Yet numeric trust scores introduce governance questions. A score can simplify policy, but opaque inputs can make denials hard to explain and may encode bad inventory data. The practical goal is not to make every decision mysterious; it is to combine context with clear minimum requirements for sensitive operations.
Keep policy brains separate from traffic enforcement at every workload
Separate decision from action. The control plane gathers identity, inventory, policy, and risk information to decide whether access is allowed. The data plane consists of systems that carry or enforce traffic, such as proxies, firewalls, and routers. In the book’s authorization model, enforcement asks a policy engine for a decision; a trust engine can contribute a risk assessment; and authoritative data stores supply current or historical facts.
Put checks near the resource. Enforcement should sit close to the workload so that trust does not accumulate behind a single gateway. Policies should describe logical relationships, such as which service role may reach another, rather than hard-coding today’s IP addresses. Store policy in version control so changes can be reviewed, explained, and compared with the configuration actually deployed.
This division reflects a broader systems principle: keep decision authority distinct from the components most exposed to untrusted input. It limits the chance that a bug in a traffic-handling process also compromises the rules it enforces. The design also creates operational dependencies. If the control plane is unavailable, a system must decide whether to deny access, rely on cached decisions, or accept some temporary risk. That failure policy deserves explicit design. Separating components improves security only when their interfaces, credentials, and fallback behavior are also carefully controlled.
Make device identity hard to copy, short-lived, and easy to revoke
Root identity in evidence. A clean, known-good software image gives a new device an initial basis for trust. A private certificate authority can issue each device an X.509 certificate, which proves possession of a key and carries identity information. A hardware security module or trusted platform module can protect the private key so it is harder to extract. TPM measurements can also help verify that a device started with an approved software configuration.
Trust decays with exposure. Inventory records connect a device identity to its role and expected properties. Provisioning systems can help authorize new certificates, ideally using more than one signal rather than trusting a single machine. The DigiNotar breach, in which attackers compromised certificate-signing infrastructure and issued fraudulent certificates, illustrates why the signing system itself must be guarded. Reimaging, patching, scanning, and rotating credentials help renew confidence.
The device lifecycle is a chain of custody problem as much as a cryptography problem. Hardware-backed keys make copying credentials harder, but they do not prove that a currently running device is uncompromised. The book usefully combines identity with freshness, inventory, and software-state evidence rather than treating a certificate as a permanent badge of safety. Each signal has limits: a freshly reimaged device can still inherit a vulnerable image, and an accurate inventory can become stale. Security depends on maintaining the whole chain, not buying one strong component.
Spend authentication friction where risk rises, not on every routine action
Let risk set the hurdle. Users should be identified separately from their devices. Passwords, one-time codes, certificates, hardware security tokens, and biometrics offer different kinds of evidence, and combining different factor types can make theft harder. A routine request may need little friction, while access to a sensitive system can require stronger proof, such as a hardware token or approval through a separate channel.
Keep authority current. Single sign-on centralizes authentication, but authorization should be checked frequently enough to revoke access when circumstances change. A long-lived application session can outlast a user’s trust or permissions. Account recovery deserves equal care: the book cites a 2012 incident in which information obtained from one online account helped an attacker convince support staff to reset access elsewhere. Recovery is part of the authentication system, not an administrative afterthought.
Adaptive authentication treats security as a scarce attention budget. Stronger checks can be reserved for actions with greater consequences, reducing the incentive to bypass controls that interrupt ordinary work. The design also recognizes that authentication methods have different failure modes: a stolen token, a phished password, and a compromised recovery channel are not equivalent risks. Risk-based systems can nevertheless frustrate users when their decisions are opaque or inconsistent. Clear explanations and dependable recovery paths are security features because they encourage people to use the intended process.
Trust software only when every transformation from source to runtime is verifiable
Secure the whole pipeline. Trust in an application must survive the journey from developer-written source through build, distribution, and execution. Signed commits can help verify who contributed code; code review provides another person’s scrutiny. But trustworthy source alone is insufficient if a compromised build system inserts malicious code. Reproducible builds address this gap by making it possible for independent parties to produce and compare identical binaries from the same inputs.
Make releases immutable. Sign artifacts or signed manifests so consumers can verify what they received, and avoid replacing a released artifact with different contents under the same version. At runtime, short-lived credentials tied to deployment can help distinguish authorized instances from abandoned or rogue ones. Isolation limits what a compromised application can reach, while active monitoring can detect unexpected behavior.
The pipeline is a supply chain: each stage inherits confidence from earlier stages while creating new opportunities for tampering. Cryptographic signatures make provenance auditable, but they cannot establish that the code is benign or that the signer acted wisely. Reproducibility adds an independent comparison, which is more persuasive than asking consumers to trust one build server. The approach also depends on careful key custody and review practices. Verification is strongest when no single person or system can silently authorize, build, sign, and deploy a change.
Encrypt and authenticate endpoint-to-endpoint, then filter both sides of each flow
Protect the conversation. Encryption hides contents; authentication verifies who sent them and whether they were altered. The book treats authenticated communication as essential and recommends encryption broadly, including inside datacenters. Mutual TLS suits many client-to-service connections, while IPsec can protect host-to-host traffic, especially in more controlled server environments. Both approaches need automation to manage certificates, keys, and policy consistently.
Reduce exposed surface. Single Packet Authorization, or SPA, lets a client send a small authenticated signal before a protected service opens a narrow, temporary firewall rule. Filtering should also occur at the host, at both endpoints of a connection, and where useful in the network. Bookended filtering means sender and receiver each enforce expected communication. If one side is misconfigured, the other can still block an unintended flow.
The layered design avoids asking one control to solve every problem. Encryption protects confidentiality and integrity, while filtering reduces which services are reachable at all. SPA adds a useful distinction between proving intent to connect and authenticating the full session: it can hide a service, but does not replace a mature protocol such as mutual TLS. Distributed enforcement can also create uneven behavior during updates. The advantage of removing a central bottleneck comes with a need for monitoring and reliable coordination across many endpoints.
Map real traffic first, then migrate policy in reversible, measured stages
Discover before denying. Start with a diagram of intended systems and gather flow data from production traffic. Group observed connections into logical relationships, then write narrowly scoped rules. Configuration management can be a practical first control plane: it can distribute firewall settings, certificates, and other expected state, though it is a stepping stone rather than an ideal permanent authorization system.
Test without breaking work. Roll out candidate rules in logging-only mode, examine what would be blocked, and enforce gradually. Google’s BeyondCorp migration used traffic simulation and required devices to show more than 99.9% eligible traffic for 30 continuous days before testing enforcement; after more than 99.99% eligible traffic during a 30-day enforcement period, devices could be considered for migration. PagerDuty also phased firewall and IPsec changes while production traffic continued.
The migration method treats security change as a reliability problem, not merely a policy-writing exercise. Logging-only deployment surfaces undocumented dependencies before they become outages, while staged enforcement limits the blast radius of mistakes. The Google thresholds demonstrate how operational evidence can support a large transition, but their exact values are not universal targets; different services have different tolerance for disruption. The deeper lesson is to make policy observable and reversible. A secure design that repeatedly breaks legitimate work will invite exceptions and eventually weaken itself.
Design for stolen identities and control-plane failure, not impossible invulnerability
Name the limits. Zero trust is not protection against every adversary. Stolen user and device credentials can still pass checks, and volumetric denial-of-service attacks can still overwhelm reachable systems. Physical coercion and social engineering can also defeat technical safeguards. Encryption may hide message contents without hiding which systems communicate, so confidentiality should not be confused with privacy.
Guard the authority. A compromised policy engine could authorize access throughout the system, making control-plane security crucial. Restrict administrative access, monitor changes, and consider requiring multiple people to approve highly sensitive actions. Also plan how quickly ongoing sessions can be revoked: request-by-request checks allow faster response than long-lived sessions. The goal is to reduce the reach and duration of compromise, detect what cannot be prevented, and keep the most sensitive assets from depending on one person or credential.
The adversarial view keeps the architecture from becoming a promise of perfect safety. Any centralized source of authority is a high-value target, even in a system designed to distrust the network. Multiple approvals can reduce unilateral risk, though they add delay and may become ritual if invoked too often. Likewise, rapid revocation improves containment but can interrupt legitimate work or depend on an available policy service. Mature design makes these trade-offs explicit, matching the strength of controls to the cost of failure rather than treating every resource as equally critical.
Analysis
The book’s central contribution is a change in where trust is placed. Traditional network security often treats the network interior as a privileged space and spends much of its effort protecting the boundary. This approach argues that modern connectivity has dissolved the conditions that once made that assumption plausible. A compromised workstation can initiate outbound connections, acquire remote commands, and use ordinary internal access to reach systems that perimeter firewalls were never designed to distinguish.
The proposed alternative is not simply more encryption or stricter firewalls. It is a coordinated operating model: authenticate users, devices, and applications; authorize the combined context of each request; encrypt communication between endpoints; and automate enforcement close to the resources being protected. The control plane interprets identity, inventory, policy, and risk, while the data plane applies decisions. This makes security less dependent on stable network geography and more compatible with cloud infrastructure, mobile work, and changing workloads.
The architecture’s strength is also its principal cost. Fine-grained access depends on accurate inventories, robust certificate management, policy maintenance, monitoring, and a dependable control plane. Automation reduces repetitive human work but concentrates authority in provisioning, signing, deployment, and policy systems. The book recognizes this concentration and argues for isolation, short-lived credentials, versioned policy, and multi-person control over critical changes. Those safeguards are not optional polish; they are part of the threat model.
A further tension lies in adaptive trust scores. They can compress many behavioral signals into practical decisions, but scores are only as defensible as their data, assumptions, and correction mechanisms. Organizations need ways to diagnose false alarms and account for missing or misleading information. The book’s implementation advice is therefore especially valuable: observe flows, test in logging mode, and move incrementally. Zero trust is best understood as a direction of travel, not a product or a claim that compromise is impossible. Its measure of success is how much unauthorized reach and dwell time it removes when prevention fails.
Review Summary
Zero Trust Networks receives mostly positive reviews, with an average rating of 4.02/5. Readers praise its comprehensive overview of zero trust concepts, though some find it too theoretical. The book is lauded for explaining the philosophy behind zero trust networks and comparing them to traditional security models. Critics note a lack of practical implementation details and suggest it may be most relevant for large organizations. Overall, it's considered a valuable introduction to an emerging security paradigm, despite some limitations in applicability.
People Also Read
Glossary
Zero trust network
Network with no implicit trustA network design that assumes communication channels may be hostile and that internal threats can exist alongside external ones. It does not use network location alone to grant trust. Users, devices, applications, and flows are authenticated and authorized, with policy informed by current system data and enforced close to the resources being accessed.
Network agent
Combined request contextAn ephemeral collection of information about the participants in a network request, typically a user, application, and device. It is assembled from authenticated identifiers and relevant data such as roles, device properties, location, and trust scores. The agent is used for authorization, not authentication, so policies can evaluate the complete context of a request.
Trust chain
Delegated trust pathA sequence through which trust granted by an operator is delegated to systems that act on the operator’s behalf. For example, an operator may authorize a provisioning system to create hosts, and that system can then provide evidence that a new host was legitimately created. The original trusted authority is the trust anchor.
Control plane
Policy and decision systemsThe part of a zero trust architecture that receives access requests, gathers relevant context, evaluates policy, and coordinates enforcement. Its services can include policy and trust engines, identity systems, and authoritative data stores. Because it can grant access across the network, it is highly privileged and requires strong isolation and protection.
Data plane
Traffic processing systemsThe systems that directly carry or enforce network traffic, including applications, proxies, firewalls, and routers. They apply the access decisions coordinated by the control plane and must operate quickly because they sit in the traffic path. The book recommends placing enforcement near endpoints so access does not become implicitly trusted behind a distant gateway.
Trust engine
Request risk assessorA component that analyzes information about a user, device, application, or request and produces a numerical assessment of risk. It can use current inventory, historical activity, explicit rules, and machine-learning models. The policy engine can then use that assessment as one input to authorization, alongside specific requirements such as role or device type.
Secure introduction
Trusted onboarding of a systemThe process of establishing an initial trust relationship with a new device or service. Because the first communication may not yet have strong credentials, secure introduction relies on an already trusted party to establish expectations and validate the new entity. The book recommends that introduction credentials be single-use, short-lived, and issued through a trusted third party.
Single Packet Authorization (SPA)
Authenticated access signalA pre-authentication method in which a client sends a small, encrypted or signed UDP packet to signal an intended connection. A valid packet can prompt a host firewall to open a narrow, temporary rule for that client. SPA helps address the first-packet problem and can conceal a service, but it does not replace full device authentication or a secure session protocol.
Bookended filtering
Filtering at both endpointsA filtering approach in which both the sender and receiver enforce which communication flows are expected. The receiving host can reject unauthorized incoming traffic, while the sending host can restrict where it is permitted to connect. This provides a second line of defense if one endpoint’s rules are loosened or misconfigured.
Reproducible builds
Independent matching software buildsA build process designed so the same source code and build inputs produce identical software artifacts across separate build environments. Independent parties can compare their output with a distributed binary to detect unexpected changes or build-system interference. Reproducibility strengthens confidence in the software supply chain, though it does not by itself prove that the source code is safe.
FAQ
What's Zero Trust Networks about?
- Security Paradigm Shift: Zero Trust Networks by Evan Gilman introduces a security model that assumes no trust in any network, internal or external, requiring verification for every access request.
- Framework for Secure Systems: The book provides a comprehensive framework for designing secure systems in untrusted environments, moving away from traditional security assumptions.
- Holistic Approach: It covers user, device, and application trust, offering strategies to manage these elements within a zero trust architecture.
Why should I read Zero Trust Networks?
- Modern Threats Addressed: The book is crucial for understanding and combating cybersecurity threats that exploit traditional perimeter defenses.
- Practical Guidance: It offers actionable design patterns and considerations for real-world systems, valuable for network engineers and security professionals.
- Expert Insights: Authors Evan Gilman and Doug Barth share their extensive experience with zero trust principles across various organizations.
What are the key takeaways of Zero Trust Networks?
- Assume Hostility: The network is always assumed to be hostile, shifting focus from perimeter security to rigorous verification of every access request.
- Dynamic Policies: Emphasizes the need for policies that adapt based on real-time data and context, rather than static rules.
- Trust Management: Introduces trust scores based on user and device behavior for nuanced access control decisions.
What is the zero trust model as defined in Zero Trust Networks?
- No Implicit Trust: Operates on the principle that no entity should be trusted by default; every access request must be authenticated and authorized.
- Continuous Verification: Requires ongoing verification of user identities and device integrity to ensure legitimate access.
- Dynamic Security Policies: Policies are dynamic, based on multiple data sources, allowing for fine-grained access control.
How does Zero Trust Networks define a zero trust network?
- Five Fundamental Assertions: The network is hostile, threats exist at all times, locality is not sufficient for trust, every access must be authenticated, and policies must be dynamic.
- Contrast with Traditional Models: Focuses on securing every interaction and minimizing trust assumptions, unlike traditional perimeter defenses.
- Holistic Security Approach: Integrates security into core operations rather than treating it as an add-on.
How does Zero Trust Networks address identity management?
- Centralized Identity Provider: Advocates for a centralized system to manage user authentication and authorization, ensuring consistent identity verification.
- Multifactor Authentication: Emphasizes the importance of MFA to strengthen identity security and reduce unauthorized access risks.
- Regular Audits and Updates: Suggests regular audits and updates to identity management practices to adapt to new threats.
What are the roles of the control plane and data plane in a zero trust network?
- Control Plane: Manages access requests, enforces policies, and maintains the network's security posture, dynamically configuring the data plane.
- Data Plane: Handles actual traffic and enforces control plane decisions, ensuring only authorized requests are processed.
- Separation of Responsibilities: Minimizes lateral movement risk, with the control plane continuously monitoring and adjusting access.
How does Zero Trust Networks suggest managing trust in users?
- Separate User and Device Trust: User trust should be managed independently from device trust due to different security implications.
- Identity Authority: Establishing a centralized identity authority is crucial for validating user identities and managing access rights.
- Adaptive Authentication: Advocates for methods that adjust based on trust scores and context, enhancing security without compromising user experience.
What are the best practices for implementing zero trust as outlined in Zero Trust Networks?
- Start with a Risk Assessment: Conduct a thorough assessment to identify critical assets and vulnerabilities, prioritizing focus areas.
- Implement Least Privilege Access: Ensures users have only necessary permissions, minimizing potential damage from compromised accounts.
- Continuous Monitoring and Response: Establish mechanisms to detect and respond to suspicious activities in real-time, maintaining network integrity.
How does Zero Trust Networks address the challenges of legacy systems?
- Authentication Proxies: Suggests using proxies to bridge legacy systems with zero trust principles, allowing secure access without infrastructure compromise.
- Incremental Adoption: Encourages gradual integration of zero trust principles with legacy systems, enhancing security measures over time.
- Secure Introduction Focus: Highlights the importance of secure introduction processes for legacy devices, ensuring authenticated and trusted connections.
What are the challenges of transitioning to a zero trust network mentioned in Zero Trust Networks?
- Cultural Resistance: Transition may face resistance from employees used to traditional security practices, requiring effective communication and training.
- Complexity of Implementation: Involves reconfiguring systems and processes, posing technical challenges that organizations must prepare for.
- Resource Allocation: Requires substantial resources, including time, personnel, and budget, necessitating careful planning to avoid operational disruption.
What are the potential attack vectors against a zero trust network as discussed in Zero Trust Networks?
- Identity Theft: Compromised user or device identities can bypass security measures, making identity protection crucial.
- Distributed Denial of Service (DDoS): Networks can still be vulnerable to DDoS attacks, requiring additional mitigation measures.
- Social Engineering: Exploits human vulnerabilities, necessitating training and awareness programs to recognize and respond to threats.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.