Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

Zero Trust Architecture

Zero trust succeeds or stalls on identity and inventory, not on network products. If you cannot say who and what is accessing each resource, including service accounts and AI agents, segmentation only moves the perimeter somewhere smaller.

Part of Security, Assurance & Governance, in the Build mode. 7 sources, checked .

of breaches involved credential abuse at some point in the attack (Verizon DBIR 2026)
39%
of breaches involved credential abuse at some point in the attack (Verizon DBIR 2026)
Verizon Business, 2026(opens in a new tab)
of organizations in IBM's 2026 breach study reported using access controls on AI models and data
40%
of organizations in IBM's 2026 breach study reported using access controls on AI models and data
IBM X-Force (on the Cost of a Data Breach Report 2026), 2026(opens in a new tab)
example zero trust architectures NIST built and documented with 24 industry collaborators (2025)
19
example zero trust architectures NIST built and documented with 24 industry collaborators (2025)
National Institute of Standards and Technology (NIST), 2025(opens in a new tab)

The problem

Stolen and misused credentials still run through most attacks. Verizon's 2026 DBIR finds credential abuse at some point in 39% of breaches, which still puts it on top, even though vulnerability exploitation now leads as the first step1. In third-party cloud environments, Verizon found that poor password practices and excessive access permissions took almost eight months to fix, and only 31% were fully remediated1. On a flat network with broad standing access, each of those becomes a wide breach.

AI adds identities that older designs never planned for. Agents and data pipelines run under service accounts that read data stores, call APIs and act without a person in the loop. IBM's 2026 study found that only 40% of organizations reported using access controls on AI models and data2. The core idea of NIST SP 800-207 applies directly: no implicit trust based on network location, and authentication and authorization of both subject and device before each session3.

The federal playbook is mature enough to borrow. OMB M-22-09 set agencies zero trust goals for the end of fiscal 2024, including phishing-resistant MFA for staff and treating every application as internet-accessible4. CISA's Zero Trust Maturity Model v2.0 scores five pillars from Traditional to Optimal5. In January 2026 NSA began publishing Zero Trust Implementation Guidelines, starting with a Discovery phase on visibility into critical data, applications, assets, services and access activity6.

Good looks like a sequenced program, not a product purchase. NIST built 19 example zero trust architectures with 24 industry collaborators and notes that every one is a custom build that starts with understanding who is accessing what resources and why7. CISA warns the path may take years5. Start with identity and inventory for the platforms your AI and data run on, then segment and tighten privileges in stages.

What you get

  1. 01

    Identity foundation for people, workloads and agents

    Enterprise identity for every person, service and AI agent, with phishing-resistant MFA for people and scoped, short-lived credentials for service accounts.

  2. 02

    Resource and access inventory

    A map of the data stores, models, APIs and pipelines on your AI and data platforms, who and what accesses each one, and the business reason for that access.

  3. 03

    Access decision and enforcement design

    Where access decisions are made and enforced for each session, and which identity, device and context signals they use, designed to NIST SP 800-207.

  4. 04

    Segmentation and least privilege

    Separation of training, inference and data environments so one compromised workload cannot reach everything, with privileges cut to what each identity needs.

  5. 05

    Maturity baseline and roadmap

    A scored baseline against the CISA Zero Trust Maturity Model pillars and a sequenced plan for moving each pillar up a stage.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • Buying a product and calling it zero trust

    Zero trust is an architecture, and NIST describes every implementation as a custom build7. A tool bought first ends up defining the design. The control: start from the access inventory and a target-state design, then choose products that fit it.

  • Forgetting non-human identities

    Service accounts and agents often hold broad standing access that nobody reviews, and excessive permissions in third-party clouds took months to fix in Verizon's data1. The control: inventory machine identities, give each an owner, issue short-lived credentials and review privileges on a schedule.

  • MFA that can be phished

    SMS codes and push prompts can be intercepted or approved by mistake. OMB M-22-09 told agencies to stop supporting SMS, voice, one-time codes and push notifications for staff and use phishing-resistant methods instead4. The control: FIDO2 or PKI-based authentication, starting with privileged and remote access.

  • Segmenting before you can see

    Segmentation designed without traffic and dependency data breaks applications and gets rolled back. NSA's implementation guidelines open with a Discovery phase that establishes visibility before later phases6. The control: baseline flows and dependencies first, then enforce in monitored stages.

Standards and rules

  • NIST SP 800-207, Zero Trust Architecture

    Primary text on csrc.nist.gov(opens in a new tab)

    The reference definition: no implicit trust based on network location, with authentication and authorization performed before each session to a resource.

  • CISA Zero Trust Maturity Model v2.0

    Primary text on cisa.gov(opens in a new tab)

    Five pillars (Identity, Devices, Networks, Applications and Workloads, Data) and three cross-cutting capabilities, each scored Traditional, Initial, Advanced or Optimal; a practical yardstick for your roadmap.

  • OMB M-22-09, Federal Zero Trust Strategy

    Primary text on whitehouse.gov(opens in a new tab)

    Binding on federal agencies and a useful model for others: enterprise identity, phishing-resistant MFA at the application layer, and applications treated as internet-accessible.

  • NSA Zero Trust Implementation Guidelines (2026)

    Primary text on nsa.gov(opens in a new tab)

    A phased series of activity-level guides that opens with a Discovery phase on visibility into critical data, applications, assets, services and access activity before later implementation phases.

How to measure it

  • Share of human accounts on phishing-resistant MFA
  • Share of service accounts and AI agents with an owner, scoped permissions and short-lived credentials
  • Share of AI and data resources behind identity-aware access with per-session checks
  • Number of standing privileged accounts, trending down
  • CISA maturity stage by pillar, reassessed each quarter

Questions buyers ask

Where should a zero trust program start?

With identity and inventory. Consolidate on enterprise identity, move privileged and remote users to phishing-resistant MFA, and map which people, services and agents access your most sensitive data and AI platforms. That map tells you where segmentation and tighter privileges will cut the most risk, and it gives you a baseline to measure progress against.

How does zero trust apply to AI agents?

Treat each agent as an identity with its own owner, scoped permissions and short-lived credentials, never as a shared service account with broad access. Check authorization on every request the agent makes to data or tools, log those decisions, and segment the environments it runs in so a manipulated agent cannot reach systems outside its task.

Do we have to get rid of our VPN?

Not on day one. The federal strategy moves users toward logging into applications rather than networks and treats every application as if it were on the internet. Most firms get there by putting identity-aware access in front of their most sensitive applications first, then retiring broad VPN access as each application moves over.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.2026 Data Breach Investigations Report(opens in a new tab)Verizon Business, May 19, 2026 · verizon.com
  2. 2.AI-powered adversaries and the enterprise risk challenge: Preparing for the new reality(opens in a new tab)IBM X-Force (on the Cost of a Data Breach Report 2026), July 29, 2026 · ibm.com
  3. 3.NIST SP 800-207: Zero Trust Architecture(opens in a new tab)National Institute of Standards and Technology (NIST), August 2020 · csrc.nist.gov
  4. 4.M-22-09: Moving the U.S. Government Toward Zero Trust Cybersecurity Principles(opens in a new tab)Executive Office of the President, Office of Management and Budget, January 26, 2022 · whitehouse.gov
  5. 5.Zero Trust Maturity Model, Version 2.0(opens in a new tab)Cybersecurity and Infrastructure Security Agency (CISA), April 2023 · cisa.gov
  6. 6.NSA Releases First in Series of Zero Trust Implementation Guidelines(opens in a new tab)National Security Agency, January 14, 2026 · nsa.gov
  7. 7.NIST Offers 19 Ways to Build Zero Trust Architectures(opens in a new tab)National Institute of Standards and Technology (NIST), June 11, 2025 · nist.gov