STRIDE Was Designed in 1999. Your Attack Surface Was Built for 2026.

Threat modelling is the one security activity that happens before anything is built, costs almost nothing to run, and can prevent more damage than any control you buy afterwards. It is also the activity most likely to be done with a framework that predates the architecture it is being applied to.

STRIDE was published in 1999. It is a genuinely good piece of thinking — a mnemonic that lets an engineer look at a data flow diagram and ask six useful questions about each element. But it was designed for a world of servers you owned, networks you could draw a boundary around, and attackers who arrived through the front door of an application you had built yourself.

That is not the world your system runs in. Your application is assembled from third-party services, your identities live in an external provider, your infrastructure is declared in code that a pipeline deploys, and a meaningful share of your attack surface belongs to vendors whose names are not on your architecture diagram. When the model does not describe the estate, the threats it produces are not the threats you face.

A note on scope. Microsoft's STRIDE is used in this article as a reference framework, not as a target. It remains one of the most influential contributions to secure design, it is still the right starting point for many teams, and the six questions it asks are as useful today as they were in 1999. Nothing here is intended to diminish that value or importance.

The point is narrower: the shape of the systems we build has changed enough that a taxonomy organised around components and a drawn boundary now needs to be complemented by something that reflects today's reality.

The numbers moved faster than the models

The 2026 Verizon Data Breach Investigations Report — its nineteenth edition, covering more than 31,000 incidents and 22,000 confirmed breaches across 145 countries — is a useful corrective, because it describes what actually happened rather than what a taxonomy expected to happen.

31%
of breaches began with vulnerability exploitation — now the top initial access vector, up from 20%.
DBIR 2026
48%
of breaches involved a third party, up from 30% — a 60% rise in a single year.
DBIR 2026
48%
of breaches featured ransomware; within system intrusions alone, 77%.
DBIR 2026
26%
of known-exploited vulnerabilities were fully remediated, down from 38%; median patch time 43 days.
DBIR 2026

Read those four together and a shape appears. Attackers are entering through software you did not write, through suppliers you do not control, and through unpatched exposure that your operating model cannot keep up with — then monetising through ransomware. Credential abuse has fallen to 13% as a first action, but it appears at some point in 39% of breaches, which tells you it is not a lesser problem; it is a later-stage one.

A threat model that cannot represent a supplier, a cloud control plane, or a sequence of actions is not wrong about these findings. It is silent on them.

What STRIDE still does well

It is worth being precise about this, because the argument is not that STRIDE is bad. It is that STRIDE is a taxonomy of what can go wrong with a component, and taxonomies have a job.

Still worth keeping

  • It gives engineers a shared vocabulary that survives contact with a design review.
  • Applied per element against a data flow diagram, it forces coverage of areas teams otherwise skip — repudiation and tampering get discussed because the checklist says so.
  • It is cheap, teachable and tool-supported, which is why it spread.
  • It scales down to a single service, which most risk-centric frameworks do not.

What it was built to assume

  • A system boundary you drew, and a diagram that stays accurate.
  • Components that are yours, or at least visible.
  • Threats that arrive as discrete events against a single element.
  • Security properties, not adversary behaviour.
  • Design-time only — the model is finished when the review ends.

What has changed, and what has not

"STRIDE is old" is often said loosely, and the detail matters — so it is worth separating three things that are usually bundled together.

The taxonomy has not changed. The six categories described in The Threats to Our Products in 1999 are still the six categories. No category has been added for cloud control planes, identity federation, software supply chains or misconfiguration. Loren Kohnfelder, who coined the acronym, has consistently described STRIDE as a taxonomy of threats to software rather than a complete methodology — a distinction worth holding on to, because a taxonomy is exactly the kind of artefact that does not age into new categories on its own.

The method and the tooling have. Microsoft folded STRIDE into the Security Development Lifecycle from 2004, and the technique was refined from a brainstorming aid into "STRIDE per element" and later "STRIDE per interaction" — changes that made enumeration more repeatable against a data flow diagram. The Microsoft Threat Modeling Tool is still shipping releases, the most recent in November 2025. Those releases are maintenance, though: bug fixes, accessibility work and file-handling safeguards rather than new threat coverage. The CMU SEI's survey of threat modeling methods makes the same distinction, noting that Microsoft no longer maintains STRIDE as a method while the tool built around it remains available.

The estate has changed most of all — and that is the actual argument. The refinement STRIDE received was about applying six categories more consistently to a diagram. It was not about whether those categories describe an environment where the boundary is set by an identity provider, where a large share of the components are vendors, and where the deployment pipeline holds production credentials.

None of this is a lonely position. Kohnfelder, who coined the acronym, describes STRIDE as a taxonomy rather than a methodology. Adam Shostack — who carried it through Microsoft's SDL, wrote Threat Modeling: Designing for Security, and in 2023 called a cloud threat model "a nice illustration of the danger of limiting oneself to STRIDE" — has made the same argument for two decades. PASTA was built explicitly against checklist-based methods; Continuous Threat Modeling exists because the workshop-as-ceremony model cannot keep pace; and the cloud threat modelling literature has converged on the gaps described here. Our contribution is not the diagnosis. It is the assembly: connecting those critiques to breach evidence and adversary behaviour in a form a team can actually run.

CyberSafeHaven's CEO and Adam Shostack holding copies of Threats: What Every Engineer Should Learn from Star Wars at OWASP AppSec Days Pacific Northwest 2023.
CyberSafeHaven's CEO with Adam Shostack at OWASP AppSec Days Pacific Northwest, Portland, 10 June 2023, with a signed copy of Threats: What Every Engineer Should Learn from Star Wars. Shostack delivered the opening keynote that day; the schedule included a book signing.

Five places the model no longer reaches

Identity became the control plane, and STRIDE treats it as one element. "Spoofing" describes impersonating a principal. It does not describe stealing an OAuth token, replaying a session cookie, consent-phishing an application into your tenant, or abusing MFA fatigue. More importantly, it does not describe the thing that now determines impact: what the compromised identity was allowed to do. A stolen credential with a standing owner role is a different threat from the same credential with read-only access to one bucket, and STRIDE has no way to say so.

Third parties are outside the diagram, so they are outside the model. Nearly half of breaches now involve a third party, across three distinct shapes: a vulnerability in a vendor's product inside your supply chain, a vendor holding your data in their environment, and a vendor with a connection into yours. None of those shapes can be drawn on a diagram that starts at your application boundary. The DBIR also notes that only 23% of third-party cloud accounts with missing or improper MFA were fully remediated, and that weak passwords and permission misconfigurations took close to eight months to half-resolve. That is a threat class, not an operational footnote.

Misconfiguration is not an actor-driven threat, and most breaches now involve one. STRIDE's six categories all presume an adversary doing something. But an over-permissive storage bucket, a public snapshot, a permissive trust policy or an exposed management endpoint is a threat that exists with no attacker present. It becomes a breach the moment someone scans for it. If your model only enumerates what an attacker could do to a component, you will never enumerate the components that should not have been reachable in the first place.

Threats arrive as sequences, and the taxonomy is a list. Ransomware appears in 48% of breaches, but ransomware is not a threat type — it is a sequence that begins somewhere, escalates somewhere, and ends in impact. STRIDE's nearest category is denial of service, which describes the last five minutes of a multi-week intrusion and none of the escalation that made it possible. A model that cannot express "and then" cannot express impact.

The estate changes weekly; the model changes annually. A pipeline deploys infrastructure on Tuesday, a team adds a SaaS integration on Thursday, and the diagram in the threat model repository was last edited before either. Combined with an average 43-day lag on known-exploited vulnerabilities, the gap between what you modelled and what you are running is now measured in months.

The DBIR test: can the model express what actually causes breaches?

The cleanest way to evaluate any threat modelling framework is to take the top observed breach patterns and ask whether the framework has a place to record them. Applied to STRIDE against the 2026 dataset, the result is uncomfortable.

Observed pattern (DBIR 2026) Where it lands in STRIDE Verdict
Exploitation of a known vulnerability as the top initial access vector (31%) Nothing directly. "Elevation of privilege" describes the effect after exploitation, not the exposed entry point or the patch debt that allowed it. Blind spot
Third-party involvement in 48% of breaches Nothing. There is no representation of a supplier, a SaaS tenant, a software dependency or a build pipeline. Blind spot
Credential abuse present in 39% of breaches at some stage "Spoofing" captures impersonation, but not token theft, session replay, consent abuse, MFA fatigue or the privilege the identity carried. Partial
Ransomware in 48% of breaches, 77% of system intrusions "Denial of service" describes the availability loss at the end, not the intrusion, escalation and exfiltration that preceded it. Partial
Only 26% of known-exploited vulnerabilities fully remediated; 43-day median Nothing. Patch latency is a property of the operating model, not a threat category. Blind spot
Misconfigured cloud storage, trust policies and permissions Nothing. Every STRIDE category assumes an actor; misconfiguration exists without one. Blind spot

Four blind spots out of six is not a tuning problem. It is a model built for a different shape of system, being asked to describe an estate it was never designed to see.

What ATT&CK gives a threat model that a taxonomy cannot

MITRE ATT&CK is often described as a threat modelling framework. It is not — it is a knowledge base of observed adversary behaviour, and that is precisely why it is useful here. A taxonomy tells you what kinds of harm exist. ATT&CK tells you what an adversary actually does, in what order, using which technique, and what telemetry that behaviour leaves behind.

The October 2026 release, ATT&CK v20, makes the point better than any argument. Enterprise now carries 15 tactics, 227 techniques, 479 sub-techniques, 178 tracked groups and 57 campaigns — along with 703 detection strategies and more than 1,700 analytics. The release itself added AI-focused and AI-enabled techniques, and new campaigns including attacks on US water utilities.

More revealing is the cloud matrix, which spans IaaS, SaaS, identity providers and office suites. It now includes techniques that no 1999 taxonomy could have anticipated: direct and indirect prompt injection, poisoned pipeline execution, cloud administration command, theft of application access tokens, and additional cloud roles. The tactics have shifted too — the cloud matrix carries stealth and defense impairment alongside the familiar credential access and lateral movement, which is a far more accurate description of how cloud intrusions actually unfold than "defense evasion" ever was.

Three things follow from this, and they are the reason ATT&CK belongs inside the modelling process rather than in the detection team's backlog:

  • It converts a category into a path. "Information disclosure" is a finding. "Exploit public-facing application → create cloud account → add additional cloud roles → discover cloud storage → exfiltrate over web service" is a plan you can defend against at five points.
  • It makes impact concrete. Each hop tells you what the adversary gains and what they can reach next. That is how you answer the only question a board asks: how far can they get, and what do they touch when they arrive?
  • It connects prevention to detection. Every technique carries detection strategies and data components. A modelled path with no detection on any hop is a known blind spot, quantified — not an assumption that monitoring will handle it.

A framework for modern applications: REACH

What follows is the approach we use at CyberSafeHaven. It keeps STRIDE's discipline for what STRIDE is good at, and replaces the parts that no longer describe reality. Five stages, each producing an artefact someone can act on.

What STRIDE covers compared with what REACH covers Left: STRIDE, six categories applied inside a data flow diagram boundary. Right: REACH, five moves covering the identity plane, third parties, pipelines, adversary chains and consequence. A strip below lists threats invisible to STRIDE but explicit in REACH. STRIDE 1999 · six categories, one boundary S Spoofing T Tampering R Repudiation I Info disclosure D Denial of service E Elevation Your data flow diagram components you own · boundary you drew threats stop at the edge Six questions per element, inside the line. Everything outside is out of scope REACH 2026 · five moves, the whole estate R Register the surface Identities, third parties, pipelines, shadow estate — not just the diagram E Enumerate entry patterns DBIR-observed vectors, each ruled in or explicitly ruled out A Amplify into ATT&CK chains Technique to technique across cloud, identity and container matrices C Cost the consequence Blast radius, crown jewels reached, detection gaps on every hop H Harden, hold, re-run Decisions per path, refreshed when the estate changes INVISIBLE TO STRIDE, EXPLICIT IN REACH token & session theft supplier compromise cloud role abuse poisoned pipeline public buckets prompt injection patch debt ransomware chains

Scroll sideways to see the full comparison

Figure 1. What STRIDE’s six categories can reach, and what REACH adds. The six questions still apply — they simply stop at the boundary of the diagram, while identities, suppliers, pipelines and the cloud control plane sit outside it.

R — Register the surface that actually exists

Model the estate, not the diagram. That means the identity graph (who and what can authenticate, with which privileges), the external dependency map (SaaS tenants, APIs, managed services, data processors), the build and deploy path, and the shadow estate nobody documented. Every entry point gets an owner and a reachability rating.

Output: an attack surface register that survives contact with reality, and a list of things you did not know you were running.

E — Enumerate observed entry patterns

Seed the model with what adversaries are measurably doing rather than with a taxonomy: vulnerability exploitation, credential abuse, phishing and pretexting, third-party compromise, misconfiguration exposure. Each pattern must be explicitly marked applicable or not applicable, with a stated reason. This changes the question from "did we think of something?" to "did we rule this out, and can we show why?"

Output: a coverage matrix where every top breach pattern is either modelled or consciously excluded.

A — Amplify each entry into an ATT&CK chain

For each applicable entry point, walk the adversary forward through ATT&CK: initial access, execution, persistence, credential access, discovery, lateral movement, collection, exfiltration, impact — using the cloud, identity and container matrices where they apply. Stop when the path reaches a crown jewel, or when a control genuinely breaks it.

Output: a set of adversary paths, each with technique identifiers, that can be reviewed, scored and attacked.

C — Cost the consequence

For every path, record what is reachable at the end, how many hops it took, what business process stops, what regulatory obligation is triggered, and — critically — which hops have detection coverage and which are dark. Impact is expressed in business terms, because "high severity" has never won a budget conversation.

Output: a ranked set of paths by blast radius and detection gap, not by category count.

H — Harden, hold and re-run

Each path resolves to a decision: prevent it at a named hop, detect it at a named hop, transfer the residual risk, or accept it with a signature. Then attach the model to change — a new integration, a new tenant, an infrastructure change, or a vulnerability entering the KEV catalogue — so the model refreshes when the estate does, not when the calendar says so.

Output: a living model with owners, evidence, and a re-run trigger.

The same system, modelled twice

Take a conventional modern application: a customer portal, an API tier, a managed database, object storage, an external identity provider, and a CI/CD pipeline deploying infrastructure as code. Here is what each approach produces.

Model Typical finding What it tells you
STRIDE "Spoofing at the portal login. Mitigate with MFA." True, useful, and complete only if the login is where the breach starts. It says nothing about the token issued afterwards, the role attached to it, or the storage bucket that role can read.
REACH + ATT&CK "Public API endpoint → exploit public-facing application (T1190) → valid cloud accounts (T1078.004) → additional cloud roles (T1098.003) → data from cloud storage (T1530) → exfiltration over web service (T1567). Three hops to customer records; detection exists on hop one only." Which control breaks the chain, which hop is dark, and what the blast radius is. The MFA recommendation survives — it now sits on a specific hop with a stated consequence instead of floating free.
REACH + ATT&CK "Pipeline has write access to production: poisoned pipeline execution → software deployment tools → additional cloud roles. No human in the loop; detection depends on commit provenance, which is not collected." A threat that never appears in a component-level taxonomy because the pipeline is not a component in the application diagram — and one that the DBIR's third-party and supply chain trend says is increasingly how breaches begin.

The difference is not that one model is more thorough. It is that one model answers "what could go wrong with this box?" and the other answers "how would someone actually get to the data, and what would stop them?"

Making it continuous, and what to measure

A better framework run once a year is still an annual framework. The operating model matters as much as the method, and it comes down to triggers and measures.

Re-run the model on four triggers: a new external integration or SaaS tenant; a change to the identity or permission model; an infrastructure change that alters reachability; and any vulnerability entering the KEV catalogue that touches your stack. Add a quarterly re-baseline against the current DBIR and ATT&CK releases, because both now update on a cadence faster than most organisations update their threat models.

Then measure things that reflect risk rather than activity. The proportion of crown-jewel paths with detection coverage on every hop. The median time from a KEV addition to that vulnerability being modelled against your estate. The proportion of third parties with a modelled entry path and a named owner. The number of paths where the only control is a policy document. None of these are comfortable numbers, which is why they are useful.

What this does not solve

Two honest limits. ATT&CK describes behaviour, not likelihood — it will not tell you which path an adversary will choose, and it is not a risk scoring model. The DBIR is a victim survey with its own sampling and reporting biases, and its percentages describe populations, not your organisation. Used together they narrow the space of things worth worrying about; they do not replace judgement about your own estate, your own adversaries, and what a day of downtime actually costs you.

What they do replace is the assumption that a framework written before cloud, before APIs became the product, and before your identity provider became your perimeter, still describes the system you are defending. It does not. The threats it cannot name are the ones that keep appearing in the breach data.

The short version

  • STRIDE's six categories are unchanged since 1999. The method and tooling around them have been maintained, but not extended for cloud, identity or supply chain.
  • The 2026 DBIR puts vulnerability exploitation (31%), third-party involvement (48%) and ransomware (48%) at the centre of breach reality; four of six top patterns have no place in the model.
  • ATT&CK v20 turns categories into paths, and paths into impact — with detection strategies attached to every hop.
  • Model the estate, not the diagram; seed from observed behaviour; chain with ATT&CK; cost the consequence; re-run on change.
  • Measure detection coverage on crown-jewel paths, not the number of threats enumerated.

Primary sources

  1. Verizon, 2026 Data Breach Investigations Report (19th edition, May 2026).
  2. MITRE, ATT&CK v20 release notes (October 2026).
  3. MITRE, ATT&CK Enterprise Cloud Matrix — IaaS, SaaS, Identity Provider and Office Suite platforms.
  4. L. Kohnfelder and P. Garg, The Threats to Our Products, Microsoft (1999) — the original STRIDE formulation.
  5. Cloud Security Alliance, MITRE ATT&CK for Cloud: A Practitioner's Guide to Detection Coverage (May 2026).
  6. CMU Software Engineering Institute, Threat Modeling: A Summary of Available Methods (2018).
  7. Microsoft, Threat Modeling Tool release notes (current release November 2025).
  8. A. Shostack, Experiences Threat Modeling at Microsoft (2008); Threats: What Every Engineer Should Learn from Star Wars, Wiley (2023).
  9. T. UcedaVélez and M. Morana, Risk Centric Threat Modeling: Process for Attack Simulation and Threat Analysis, Wiley (2015).
  10. I. Tarandach and M. Coles, Continuous Threat Modeling handbook; Threat Modeling: A Practical Guide for Development Teams, O'Reilly (2020).
  11. OWASP, AppSec Days Pacific Northwest 2023 conference schedule.

Figures quoted are from the 2026 DBIR reporting dataset, covering incidents between November 2024 and October 2025, and from ATT&CK v20 as released in October 2026.

Want your threat model to describe your actual estate?

We run threat modelling as an engineering exercise — attack surface first, adversary paths second, decisions with owners third. It works whether you are designing something new or trying to understand what you already run.