Insights · Security Architecture
Remove the DMZ: a practical guide to zero trust architecture
The DMZ grants trust by location, and that fails in the cloud. Zero trust authenticates and encrypts every request, per resource.
NIST’s definition in SP 800-207 is the anchor: zero trust means “no implicit trust granted to assets or user accounts based solely on their physical or network location,” and — the part that matters for this argument — “zero trust focuses on protecting resources (assets, services, workflows, network accounts, etc.), not network segments.” Read that twice and the DMZ question answers itself: a security model organized around network segments cannot protect resources in an estate where the segments are someone else’s hypervisor and the workloads move hourly.
Why the DMZ fails in cloud environments
The three-tier perimeter — internet, DMZ, intranet — encodes one bet: that the interior is trustworthy because it is interior. That bet fails four ways in a modern estate:
- VPN concentration. The VPN grants network-level presence, and network presence is the keys to the flat kingdom. One phished credential or stolen client certificate converts to estate-wide reach.
- Flat internal networks. Once inside, an attacker moves laterally. MITRE ATT&CK gives this its own tactic — Lateral Movement (TA0008): “the adversary is trying to move through your environment,” entering and controlling remote systems, often pivoting through multiple systems and accounts. Every quarter’s intrusion report is the same story: the perimeter held; the interior was a hallway.
- The perimeter dissolved first. SaaS, cloud workloads, partner integrations and developer laptops already live outside the DMZ. The interior stopped being an interior years ago; the policy model just didn’t notice.
- East-west traffic is unauthenticated. In the classic model, traffic between internal systems is encrypted nowhere and authorized nowhere — the compromise of one workload compromises every conversation it participates in.
A 2025 systematic literature review of zero trust architecture research (Gambo & Almulhem, arXiv) traces a decade of work converging on the same operating principle — “never trust, always verify,” continuous authentication, conditional access, dynamic trust evaluation — precisely because perimeter-based control measurably failed in distributed estates.
What zero trust architecture actually is
Zero trust is not a product purchase; it is an architectural decision about where policy lives. SP 800-207’s model separates policy decision from policy enforcement: a policy engine evaluates each request against identity, device posture, and context, and enforcement points apply the verdict at the resource. The UK NCSC’s design principles compress the posture to one sentence: “inherent trust in the network is removed, the network is assumed hostile, and each request is verified based on an access policy” — with trust in no particular network (corporate, cloud, or home) and secure transport for all communications.
The CISA Zero Trust Maturity Model v2.0 organizes the journey into five pillars — Identity, Devices, Networks, Applications & Workloads, and Data — crossed by three capabilities (Visibility and Analytics, Automation and Orchestration, Governance) and four maturity stages from traditional to optimal. The DoD’s Zero Trust Strategy set the same destination at national-defence scale, organized into seven pillars. Different taxonomies, same destination: per-request authorization anchored in identity, not address ranges.
The east-west decision: default-strict mTLS
The interior conversation is where the model gets tested. NIST’s SP 800-207A extends zero trust into cloud-native applications with the relevant tenet: “remove the implicit trust in users, services, and devices based only on their network location” — enforced through service identity infrastructure (SPIFFE-class), API gateways, and sidecar proxies that authenticate every service-to-service call. In practice that means:
- Every workload has a cryptographic identity — a certificate or SVID bound to the workload, not to an IP address.
- Mutual TLS by default, not by exception. Unauthenticated east-west traffic fails closed. Exceptions are recorded, expiring, and reviewed.
- Authorization is per-call policy — which identity may call which endpoint for which operation — not per-firewall-zone.
On the E-Tax modernization, default-strict mTLS for east-west traffic was one of the programme’s first architecture decision records — written before the first service existed, so that every later service inherited the posture instead of debating it. That sequencing is the difference between zero trust as architecture and zero trust as an aspiration.
The migration: what happens to the DMZ
Google’s BeyondCorp papers are the canonical migration narrative. The 2014 opening move is unambiguous: “We are removing the requirement for a privileged intranet and moving our corporate applications to the Internet” (Ward & Beyer, ;login:). The follow-up paper on migrating to BeyondCorp is the practical companion — phased movement off the VPN while keeping users productive, access proxies in front of applications, trust tiers computed from device inventory, and access policy evaluated per request. The migration sequence generalizes:
- Map flows. You cannot remove trust you cannot see. Discover which identities talk to which resources, on which paths, and enumerate the implicit grants your firewall rules encode.
- Identity-first for people. Phishing-resistant MFA, device posture in the decision, and per-application access proxies replace network-level VPN presence. The VPN’s residual role shrinks to nothing as the last apps move behind identity-aware access.
- Identity-first for machines. Workload identity and default-strict mTLS for east-west, as above.
- Segment to contain failure. Microsegmentation and default-deny network policy make any single foothold a local event instead of a domain-wide one.
- Shrink, then dissolve, the DMZ. What remains of the DMZ is an identity-aware proxy front door — the enforcement point of the new model. When access no longer depends on where a request comes from, the zone stops doing security work and can be retired.
- Govern continuously. Zero trust is a maturity loop, not a project: CISA’s stages exist because visibility, automation and governance keep running after the architecture lands.
The companion argument on this site — zero trust is an architecture decision, not a product purchase — covers the governance side: record the posture as decisions, and the exceptions as time-boxed deviations. The Kubernetes hardening guide covers the workload controls (non-root containers, default-deny network policies, no standing SSH) that make zero trust enforceable at the platform layer.
What zero trust does not fix
Zero trust assumes the endpoint may be hostile and the network is — but it still requires working identity infrastructure, disciplined secrets management (vaults and workload identity), and patched platforms. It raises the cost of every step an attacker takes; it does not abolish them. And identity is now the perimeter: if your IdP is the single door, its hardening, its recovery paths, and its own administrators are the crown jewels — protect them like the DMZ you just removed, minus the implicit trust.
Assess your zero trust posture
Sources
- NIST SP 800-207, “Zero Trust Architecture” (Rose, Borchert, Mitchell, Connelly; 2020)
- NIST SP 800-207A, “A Zero Trust Architecture Model for Access Control in Cloud-Native Applications in Multi-Cloud Environments” (Chandramouli & Butcher, 2023)
- CISA, “Zero Trust Maturity Model Version 2.0” (April 2023)
- U.S. Department of Defense, “DoD Zero Trust Strategy” (November 2022)
- NSA, “Embracing a Zero Trust Security Model” guidance announcement (2021)
- UK NCSC, “Zero trust architecture design principles” (v1.1)
- R. Ward & B. Beyer, “BeyondCorp: A New Approach to Enterprise Security”, ;login: 39(6), USENIX, 2014
- Peck, Beyer, Beske & Saltonstall, “Migrating to BeyondCorp: Maintaining Productivity While Improving Security”, ;login: 42(2), USENIX, 2017
- MITRE ATT&CK, “Lateral Movement” (TA0008)
- Gambo & Almulhem, “Zero Trust Architecture: A Systematic Literature Review” (arXiv, 2025)
- Weinberg & Cohen, “Zero Trust Implementation in the Emerging Technologies Era: Survey” (arXiv, 2024)