Host firewalls and network firewalls overlap just enough that organisations either duplicate every rule in both, creating two rule bases nobody can review, or assume one covers for the other and leave east-west traffic inside a zone completely open. The two controls see different traffic, fail in different ways and are usually owned by different teams, so the useful question is which flows each should own and who maintains the rules. This guide takes that narrow angle: east-west segmentation on hosts versus perimeter and zone filtering on network devices, the threat model for each, how to split rule ownership between endpoint, server and network teams, and how to measure that the layers are working together. It is for infrastructure, network and endpoint security engineers and the architects who set segmentation policy. Platform configuration is covered in the Host-Based Firewall Guide, the Windows Defender Firewall Guide and the Linux Firewall Guide.
Key takeaways
- Network firewalls filter between zones, host firewalls filter within them. Two laptops on the same VLAN or two servers in the same subnet never cross a network firewall, so only the host can stop lateral movement between them.
- Give each layer different questions. The network layer answers "may this zone talk to that zone"; the host layer answers "which peers and management hosts may reach this service on this machine". Copying zone rules onto hosts adds work without adding protection.
- Ownership follows the rule's purpose, not the device. Network teams own zone policy, platform teams own host baselines, and application owners own the narrow service exceptions, all through one request process.
- The host layer can be switched off from the host. A local administrator or malware with admin rights can disable a host firewall (ATT&CK T1686), so the network layer must still contain a compromised host's zone, and changes to host policy must be logged.
- Devices that cannot run an agent stay network-only. CISA's microsegmentation guidance notes that OT, IoT and legacy systems often cannot run host enforcement, so their protection must come from network zones.
The short answer
Use network firewalls for perimeter filtering and for separating zones of different trust (user, server, management, OT, guest, third party). Use host firewalls for default deny inbound on every endpoint and server, with narrow allows for the management hosts and peers that genuinely need access. Do not try to make either layer do the other's job: network devices cannot see same-subnet traffic, and thousands of host rule sets are a poor place to express zone policy. NIST SP 800-41 Rev. 1 describes host-based firewalls as giving more granular protection than network firewalls and recommends that personal firewall management be centralised; NIST SP 800-53 Rev. 5 names the host layer separately as SC-7(12) Host-based Protection, an enhancement of SC-7 Boundary Protection.
What each sees
| Dimension | Host firewall | Network firewall (perimeter or internal segmentation) |
|---|---|---|
| Enforcement point | The operating system of each host | A routed hop between networks or zones |
| Traffic seen | Every connection to and from that host, including same-subnet peers | Only traffic that crosses it; nothing between hosts in the same zone |
| Context available | Process or executable path, local user, network profile (location) | Zone, address, port, application identification, user mapping |
| Follows the device | Yes, including on home and public networks | No |
| Devices covered | Only those with a supported OS and managed policy | Everything behind it, including OT, IoT, printers and appliances |
| Tamper resistance | Weaker: a local administrator can change it | Stronger: separate device and separate administrators |
| Scale of rule base | One baseline applied to thousands of hosts, plus exceptions | Fewer devices, larger and more varied rule bases |
| Typical owner | Endpoint or server platform team | Network or network security team |
| Logging | Dropped and allowed connection logs per host, rule change events | Central session and threat logs |
Threat model for each
Host firewall
The host firewall assumes the attacker is already inside a zone, on a peer machine, and is trying to move sideways: SMB and admin shares (T1021.002), Remote Desktop (T1021.001), Windows Remote Management (T1021.006), copying tools between hosts (T1570) and scanning for listening services (T1046). It also protects laptops on untrusted networks, where no corporate network firewall exists. Its weakness is that it runs on the machine it protects: an attacker with administrative rights can disable or modify it (T1686 Disable or Modify System Firewall).
Network firewall
The network firewall assumes the attacker is outside a zone and must cross a boundary: from the internet inbound, from a user zone to a server zone, from IT to OT, from a third-party connection into the estate. It is independent of the hosts it protects, so a compromised host cannot reconfigure it, and it covers devices that run no agent. Its weakness is granularity: inside a zone it sees nothing, and zones tend to grow because splitting them is disruptive.
Architecture in words
- Perimeter: internet edge firewalls allow inbound only to published services and control outbound by zone.
- Zones: internal segmentation firewalls or routed ACLs separate user, server, management, OT, guest and partner zones with deny by default between them. Zone rules are coarse: "user zone may reach the web tier on TCP 443".
- Hosts: every Windows, macOS and Linux host runs a centrally managed firewall that denies inbound by default. Workstations allow almost nothing inbound; servers allow their service ports from the zones that need them, and management ports only from jump hosts or privileged access workstations.
- Management plane: host rules come from Group Policy, an MDM configuration service provider, configuration profiles or configuration management; network rules come from the firewall manager. Both pull from one request and approval workflow.
- Telemetry: host drop logs and rule change events and network session logs feed the same SIEM so that a flow can be followed across both layers.
CISA's Microsegmentation in Zero Trust Part One (July 2025) sets out the same pattern in its own terms: network-based enforcement reuses existing infrastructure and supports legacy systems but has limited visibility into endpoints and application workflows, while host-based enforcement integrates better with identities and works for remote devices but needs an agent on every endpoint. It recommends centralising policy management and visibility regardless of where policies are enforced. See the Network Segmentation Guide.
Who owns the rules
Duplicate effort mostly comes from unclear ownership: two teams each writing a rule for the same flow, or neither writing it because each assumes the other did. A workable split:
| Rule type | Layer | Owner | Approver | Review cycle |
|---|---|---|---|---|
| Internet inbound and outbound by zone | Network | Network security team | Security architecture | Quarterly |
| Zone to zone access | Network | Network security team | Owners of both zones | Quarterly or on zone change |
| Workstation baseline (deny inbound, peer blocks) | Host | Endpoint platform team | Security architecture | Twice a year |
| Server baseline (deny inbound, management from jump hosts) | Host | Server platform team | Security architecture | Twice a year |
| Application service exceptions on hosts | Host | Application owner, implemented by platform team | Security, through the change process | On application change and annually |
| Management access paths (RDP, SSH, WinRM) | Both | Privileged access owner | Security architecture | Quarterly |
One request form should capture the flow once and route it: zone crossings to the network team, host exceptions to the platform team. The Firewall Rule Request Template can be extended with a "layer" field for this.
Avoiding duplicate effort
- Do not mirror zone rules on hosts: if the network firewall already stops the user zone reaching the database zone, the database host only needs to allow its port from the application tier subnet. A full copy of zone policy on every host doubles change work.
- Use groups, not addresses: host baselines should refer to named groups (management hosts, jump hosts, application tier) maintained in one place, so a new jump host is one change, not thousands.
- Baseline once, exception narrowly: most hosts share a baseline. Exceptions belong to the application and are removed with it.
- Decide local merge deliberately: Microsoft documents AllowLocalPolicyMerge per profile, which controls whether local administrators can add rules on top of GPO or MDM rules. Disabling merge prevents shadow rule sets but means every application needing inbound access must be deployed through central rules.
- Keep host rules for what only hosts can do: process-scoped rules, peer blocks inside a zone and location-aware profiles. Leave inspection, IPS and application identification at the network layer.
Failure modes
Host layer
- Disabled or default: firewalls turned off during troubleshooting and never turned back on, or left at defaults that allow workstation-to-workstation SMB and RDP.
- Local rule sprawl: application installers and user prompts add allow rules that no central policy knows about. Microsoft notes that a non-administrator answering the first-run prompt always gets block rules, while administrators can create allows.
- Linux gaps: server images with no rules loaded, and container platforms that write their own rules underneath the host policy.
- Tampering: attackers with admin rights stop the service or add an allow rule; without rule change logging nobody notices.
Network layer
- Flat zones: a single large user or server VLAN where the network firewall sees nothing of lateral movement.
- Any-any between internal zones: temporary rules for migrations that remain years later.
- Stale objects: address objects for decommissioned servers that are now reused by different systems.
- Assumed coverage of remote devices: laptops off site are protected by no network firewall at all.
Where identity fits
Segmentation decisions increasingly use identity rather than address. Host firewalls on Windows can scope rules to authenticated connections through IPsec connection security rules, and to tagged applications through App Control AppID tags, which Microsoft documents as a way to target processes without relying on paths. Management access paths should be tied to privileged access workstations and named admin accounts, not to whole subnets. The automation that pushes policy is itself a privileged non-human identity: Group Policy editing rights, MDM service principals, configuration management credentials and firewall manager API tokens can rewrite segmentation across the estate, so vault them, scope them and log their use. See the identity segmentation vs network segmentation comparison.
Decision criteria
| Situation | Primary layer | Supporting layer |
|---|---|---|
| Laptops on home and public networks | Host | None on the network side; SSE and ZTNA for access |
| Workstation to workstation in one office VLAN | Host | Private VLAN or switch ACLs where supported |
| User zone to server zone | Network | Host allows by source subnet on servers |
| Server to server in the same tier | Host (or microsegmentation agent) | Network for tier boundaries |
| OT, IoT, printers, appliances | Network | None: no agent possible |
| Internet inbound to published services | Network | Host allows only the published port |
| Cloud workloads | Cloud security groups and firewalls | Host firewall for defence in depth |
Small organisations: one perimeter firewall plus a strong host baseline on every device usually gives more protection than investing in internal segmentation firewalls. Large organisations: need both, a microsegmentation platform where server-to-server policy grows beyond what native host firewalls can manage, and a single policy model across layers; see the Microsegmentation Buyers Guide.
Worked scenario
Regional hospital group, 4,000 endpoints, flat clinical network. A ransomware exercise shows that a compromised reception PC can reach every other workstation on TCP 445 and 3389 and most servers on TCP 135 and 5985. The network team proposes splitting the clinical VLAN into dozens of zones; the estimate is eighteen months because of medical devices. Decision: the endpoint team deploys a workstation baseline through central policy that blocks inbound SMB, RDP, WinRM and RPC from peers and allows them only from the management server group, in audit mode for two weeks, then enforced by ring. In parallel the network team creates one zone for medical devices that cannot run agents, with strict rules to the clinical systems they use. Rule ownership is written down: zone rules with the network team, host baselines with the endpoint team, exceptions requested by application owners through one form. The Clinical Network Segmentation Guide covers the device side in depth.
Implementing both layers
- Map flows: collect host firewall logs in allow-and-log mode and network session logs for two to four weeks to see who really talks to whom. CISA's guidance stresses identifying dependencies with stakeholder input before setting policy.
- Write the layer split: list zones and the flows between them for the network team; list baselines and exception groups for the platform teams.
- Audit mode first: deploy host baselines with logging but without blocking where the platform supports it, or to a small pilot ring; on the network side, add rules with logging above existing permissive rules.
- Enforce by ring: pilot, early adopters, broad deployment, sensitive systems last. Watch blocked connection counts and help desk tickets per ring.
- Tighten zone rules: remove internal any-any rules once host enforcement is in place and flows are known.
- Roll back cleanly: keep the previous host policy version ready to re-apply in one change, and keep removed network rules disabled for a defined period before deletion.
Validate the result with the Network Segmentation Testing Guide and review rules with the Firewall Rule Review Checklist.
Detection and measurement
- Host firewall coverage: percentage of managed hosts with the firewall enabled on all profiles and the current baseline applied (target 100 percent of supported hosts).
- Peer connection attempts blocked: host drops for TCP 445, 3389, 5985 and 135 from non-management sources; spikes suggest scanning or lateral movement.
- Host rule change events: on Windows, events 4946 to 4950 record firewall rule additions, changes, deletions and setting changes; alert on changes not made by the management system.
- Local rule count: number of locally created rules per host, trending down where merge is allowed.
- Internal any-any rules: count on network firewalls between internal zones, trending to zero.
- Duplicate rules across layers: flows with an identical rule on both layers, reviewed during rule recertification.
Framework mapping
| Framework | Reference | Host firewall | Network firewall |
|---|---|---|---|
| NIST SP 800-53 Rev. 5 | SC-7 Boundary Protection; SC-7(5) Deny by Default, Allow by Exception | Default deny inbound per host | Default deny between zones and at the perimeter |
| NIST SP 800-53 Rev. 5 | SC-7(12) Host-based Protection | Direct fit | Not applicable |
| NIST SP 800-53 Rev. 5 | AC-4 Information Flow Enforcement; CM-7 Least Functionality | Only required services reachable | Only required flows between zones |
| NIST CSF 2.0 | PR.IR-01; PR.PS-01; DE.CM-01 | Baselines under configuration management, host logs monitored | Zone protection and network monitoring |
| CIS Controls v8.1 | 4.4 and 4.5 host-based firewalls on servers and end-user devices; 13.4 traffic filtering between network segments | 4.4 and 4.5 | 13.4 |
| MITRE D3FEND | D3-ITF Inbound Traffic Filtering; D3-NI Network Isolation | D3-ITF on the host | D3-NI and D3-ITF at zone boundaries |
| CISA TIC 3.0 | Networking PEP: Network Segmentation, Microsegmentation, Host Containment | Microsegmentation, Host Containment | Network Segmentation |
Common failure patterns
- Two rule bases for the same flow: every network rule copied to hosts, so each change needs two teams and the copies drift apart.
- "The firewall handles it": host firewalls disabled because a network firewall exists, leaving same-zone lateral movement wide open.
- No owner for exceptions: host allow rules for applications long retired, with no record of who asked for them.
- Management access by subnet: RDP or SSH allowed from the whole IT VLAN rather than from jump hosts, so one compromised admin laptop reaches everything.
- Silent tampering: no alerting on host firewall rule changes, so an attacker's allow rule persists.
- Unmanaged devices forgotten: agentless devices left in user zones without network rules because the plan assumed host enforcement everywhere.
Practitioner checklist
- Write down which flows each layer owns and publish it to network, endpoint, server and application teams.
- Enforce default deny inbound on every supported host through central policy, and verify it daily.
- Block peer SMB, RDP, WinRM and RPC on workstations, allowing only named management groups.
- Restrict management ports on servers to jump hosts or privileged access workstations.
- Decide and document the local policy merge setting per profile.
- Place agentless devices in dedicated network zones with explicit rules.
- Alert on host firewall rule and setting changes not made by the management system.
- Run one rule request process that routes each flow to the right layer and owner.
- Recertify rules on both layers on a fixed cycle, removing duplicates and expired exceptions.
Standards and references
- NIST: SP 800-41 Rev. 1, Guidelines on Firewalls and Firewall Policy
- NIST: SP 800-53 Rev. 5, Security and Privacy Controls
- CISA: Microsegmentation in Zero Trust, Part One: Introduction and Planning
- CISA: TIC 3.0 Security Capabilities Catalog
- Microsoft Learn: Windows Firewall rules
- CIS: Critical Security Controls v8.1
- NIST: SP 800-207, Zero Trust Architecture
- MITRE ATT&CK: Disable or Modify System Firewall
Related NHI Mgmt Group resources: Firewall Buyer's Guide, Microsegmentation vs ZTNA, T1686 Disable or Modify System Firewall, Next Generation Firewall Health Check.