Network Management Research Group Y. Ahn, Ed.
Internet-Draft J. Jeong, Ed.
Intended status: Informational Sungkyunkwan University
Expires: 6 May 2027 Y. Kim
Soongsil University
2 November 2026
An Agentic AI Security Framework for Security Management in Mobile
Networks
draft-ahn-nmrg-5g-security-i2nsf-framework-03
Abstract
This document proposes an Agentic AI Security Framework for security
management in mobile networks. A Security Application Function (SAF)
Agent translates a network administrator's security intent to hold
security objectives into a high-level security policy. A Security
Control Function (SCF) Agent receives the high-level policy and
translates it into a low-level security policy that will be enforced
by distributed Network Security Functions (NSFs) in mobile networks.
The SCF Agent sends the low-level security policy to an apporiate
NSF. Policy generation, monitoring, and continuity across a data
path are integrated into a closed-loop control with an Observe-
Reason-Act-Verify process. The proposed framework can be applied to
enterprise, campus, cloud, and edge networks along with existing
network-management mechanisms. This document describes the
architecture, a policy-continuity example, standard integration
points, and security considerations.
Status of This Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF). Note that other groups may also distribute
working documents as Internet-Drafts. The list of current Internet-
Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
This Internet-Draft will expire on 6 May 2027.
Ahn, et al. Expires 6 May 2027 [Page 1]
Internet-Draft Agentic AI Security Framework November 2026
Copyright Notice
Copyright (c) 2026 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents (https://trustee.ietf.org/
license-info) in effect on the date of publication of this document.
Please review these documents carefully, as they describe your rights
and restrictions with respect to this document.
Table of Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 3
2. Terminology . . . . . . . . . . . . . . . . . . . . . . . . . 4
3. Agentic AI Architecture for Network Security Management . . . 4
3.1. SAF Agent: Goal Interpretation . . . . . . . . . . . . . 5
3.2. SCF Agent: Policy Planning and Verification . . . . . . . 6
3.3. Context Exchange and Network Control . . . . . . . . . . 6
3.4. Distributed NSFs: Policy Execution . . . . . . . . . . . 7
3.5. Observation, Analytics, and Policy Memory . . . . . . . . 7
4. Agentic Security Management Procedures . . . . . . . . . . . 8
4.1. Policy Generation, Delivery, and Continuous
Verification . . . . . . . . . . . . . . . . . . . . . . 8
4.2. Policy Continuity during an Enforcement-Point
Transition . . . . . . . . . . . . . . . . . . . . . . . 9
4.3. Policy State and Failure Handling . . . . . . . . . . . . 12
4.4. Handling an Unplanned Path Change . . . . . . . . . . . . 12
4.5. Application Example: Mobility-Driven Enforcement
Transition . . . . . . . . . . . . . . . . . . . . . . . 13
5. Multi-Domain Coordination . . . . . . . . . . . . . . . . . . 15
6. Security Considerations . . . . . . . . . . . . . . . . . . . 16
7. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 17
8. IANA Considerations . . . . . . . . . . . . . . . . . . . . . 17
9. References . . . . . . . . . . . . . . . . . . . . . . . . . 17
9.1. Normative References . . . . . . . . . . . . . . . . . . 17
9.2. Informative References . . . . . . . . . . . . . . . . . 17
Appendix A. Changes from
draft-ahn-nmrg-5g-security-i2nsf-framework-02 . . . . . . 18
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . . . 19
Ahn, et al. Expires 6 May 2027 [Page 2]
Internet-Draft Agentic AI Security Framework November 2026
1. Introduction
Distributed networks require security objectives to remain effective
as topology, traffic paths, workloads, and enforcement resources
change. A policy that protects traffic at one location can become
ineffective when the traffic moves to another path or administrative
domain. Maintaining the intended outcome therefore requires
coordination between policy interpretation, enforcement placement,
and observation of the resulting behavior.
An Agentic AI Security Framework in this document uses an intent,
intent fulfillment, and intent assurance defined in Intent-Based
Networking (IBN) in [RFC9315]. This framework develops the IBN
approach into an agentic AI security management system.
In the proposed framework, a Security Application Function (SAF)
Agent translates a network administrator's security intent to hold
security objectives into the corresponding high-level security
policy. A Security Control Function (SCF) Agent receives the high-
level policy and translates it into the corresponding low-level
security policy. This low-level security policy will be enforced by
distributed Network Security Functions (NSFs) in mobile networks
(e.g., 5G networks and WiFi networks). The SCF Agent sends the low-
level security policy to an apporiate NSF to provision the required
security service. Thus, policy generation, monitoring, and
continuity across a data path in the mobile networks are integrated
into a closed-control loop with an Observe-Reason-Act-Verify process
to satisfy the give security intent.
The framework in this document integrates policy generation,
distributed enforcement, monitoring, and policy continuity across
enterprise, campus, cloud, and edge networks. Its enforcement-point
transition in this document autonomously addresses configuration
changes caused by workload relocation, failover, rerouting, or a
change in network attachment in the mobile networks.
This document clarifies the progression from a security intent to
security enforcement, monitoring, and policy continuity with NSFs in
the mobile networks. It details the roles of the SAF and SCF into AI
agents. These AI agents receive security objectives, evaluate
context, validate outcomes of the security policy, and reconfigures
the policy in a closed-loop control. The sequence of security
enforcement in the mobile networks is autonomously adapted according
to the workload and path changes in the target mobile networks.
Agentic AI empowers persistent security objective with context-
sensitive reasoning in the proposed framework. The AI agents observe
the environment, reason permitted actions, act accordingly through
Ahn, et al. Expires 6 May 2027 [Page 3]
Internet-Draft Agentic AI Security Framework November 2026
authorized tools, and validate the results of the security
operations. The framework can revise an unsuccessful operations
while preserving the administrator's security intent and the
constraints in the mobile networks in terms of the capabilities and
resources of the available NSFs. Note that a model-generated
recommendation is validated before execution in this framework, but
the use of AI does not authorize a security policy without any report
to the network administrator.
Therefore, this document proposes an architecture of an Agentic AI
Security Framework along with an operational example. It focuses on
mobile-network-side security enforcement such as traffic filtering
and access control in the target mobile networks. Application
traffic can be protected by the proposed framework with the available
NSFs with security capabilities. However, device-specific
application configuration and unrestricted control over routing are
out of the scope of this document.
2. Terminology
* Intent: A declarative expression of desired outcomes, using the
intent definition in [RFC9315].
* SAF Agent: The security-application agent that interprets intent
and maintains the objective and constraints.
* SCF Agent: The security-control agent that plans policy actions,
coordinates execution, and verifies results.
* NSF: A Network Security Function that enforces approved filtering
or access-control policies and provides telemetry.
* Policy Context: The policy revision, protected traffic scope,
enforcement bindings, and relevant applied state.
* Enforcement Point: The NSF instance and traffic-path location at
which a policy is applied.
3. Agentic AI Architecture for Network Security Management
The architecture separates objective interpretation, security
planning, context exchange, and distributed enforcement. Figure 1
shows these logical relationships. Interface A carries the
objective, high-level policy, and outcomes between agents; interface
B supplies authorized context and change coordination; interface C
manages NSFs and returns telemetry. These labels identify
responsibilities rather than defining new wire protocols.
Ahn, et al. Expires 6 May 2027 [Page 4]
Internet-Draft Agentic AI Security Framework November 2026
The management, modeling, telemetry, and domain-integration
mechanisms cited below are implementation options rather than
prescribed dependencies. Agent coordination needs agreed definitions
for objective scope, action representation, readiness, verification,
and failure reporting. The architecture identifies these
relationships without defining a protocol-specific implementation
profile.
+--------------+ Intent / feedback +------------------------+
| User / Admin |----------------------| SAF Agent |
+--------------+ | Objective owner |
| Application integration|
+------------+-----------+
|
Proposed agent interface A |
High-level policy / outcome|
|
+------------v-----------+
| SCF Agent |
| Translate / plan / |
| verify / recover |
+------------+-----------+
|
NSF management interface C |
Low-level policy / telemetry
+--------------+-------+
| |
+------v------+ +------v------+
| Edge / NSF1 | | Edge / NSF2 |
+-------------+ +-------------+
SCF Agent ----- interface B ----- Context / controller adapter
| authorized context / coordination
+---- inventory maps traffic scope to NSF endpoints
Controllers retain routing, workload and path-change authority.
Figure 1: Agentic AI Network Security Management Architecture
3.1. SAF Agent: Goal Interpretation
The SAF Agent receives administrator intent and identifies the
protected user, workload, service, or traffic scope. AI-assisted
interpretation produces a high-level security policy describing the
desired outcome, constraints, priority, and validity. Ambiguous
requests require clarification, and unsupported or unauthorized
objectives are reported rather than silently approximated.
Ahn, et al. Expires 6 May 2027 [Page 5]
Internet-Draft Agentic AI Security Framework November 2026
The SAF Agent retains the objective independently of the current
device or path. It uses SCF feedback to distinguish satisfied,
degraded, unknown, and pending outcomes.
During an enforcement-point transition, it preserves the security
constraints and decides whether a proposed alternative or recovery
action remains within the original authorization. A changed location
does not by itself justify weakening protection.
3.2. SCF Agent: Policy Planning and Verification
The SCF Agent combines the high-level policy with current topology,
traffic bindings, NSF capabilities, and observed enforcement state.
AI-assisted reasoning supports selection of a target realization or
recovery plan. Deterministic checks validate the proposed rules,
their scope, and their compatibility with operator constraints before
an authorized executor installs them.
Planning can select an existing enforcement point, prepare a target,
or retain a reachable protected path. The SCF Agent records the
expected outcome and the evidence needed to verify it.
Installation, activation, and post-change verification remain
separate results. If the observed result does not meet the
objective, the agent can revise the plan within a bounded action
budget or request operator intervention.
The proposed architecture does not depend on a particular model or
prompting mechanism. An implementation must identify where AI
contributes interpretation or planning and how generated proposals
are checked. The agent's persistent objective and evidence-driven
replanning are the relevant behavior; a fixed rule-copy operation
alone does not provide the full decision and assurance loop.
3.3. Context Exchange and Network Control
Inventory, topology services, telemetry collectors, and existing
network controllers supply authorized observations. The SCF Agent
maps a protected traffic scope and candidate location to the
appropriate NSF endpoint and capability set. A generic topology
model such as RFC 8345 [RFC8345] can provide a basis for representing
network resources; policy bindings and technology-specific details
still need a deployment-specific model.
The security agent does not acquire ownership of routing merely
because it can configure an NSF. A path or workload change remains
under the authority of the relevant controller or orchestrator.
Ahn, et al. Expires 6 May 2027 [Page 6]
Internet-Draft Agentic AI Security Framework November 2026
Where that controller supports advance notification and preparation
acknowledgment, the agents may coordinate target readiness with the
pending transition. Otherwise, target preparation is opportunistic
and a continuity policy must cover late or unexpected changes.
The context exchange should distinguish a candidate path from a
selected or active path. Readiness prepared for one candidate cannot
establish protection on a different destination. When observations
are incomplete, the SCF Agent records the uncertainty and obtains
updated context or follows an authorized fallback rather than
claiming that the original plan remains valid.
3.4. Distributed NSFs: Policy Execution
NSFs execute approved filtering or access-control rules and return
their policy revision, binding, and operational status. They may
reside on routers, security appliances, virtual machines, containers,
or service edges. The deployment must establish the traffic path
through the selected function; simply running an NSF beside a
workload does not ensure enforcement.
The SCF Agent maintains desired policy and revision history. NSF
readback supplies evidence of the applied realization and, where
supported, transferable runtime state. Different targets can expose
different capabilities, so migration may require translation instead
of a byte-for-byte rule copy. A target that cannot preserve the
requested behavior is rejected or escalated.
YANG [RFC7950] can model supported configuration and state. NETCONF
[RFC6241] or RESTCONF [RFC8040] can provide NSF management interfaces
when suitable models and capabilities are implemented. These choices
do not make the proposed SAF-SCF interface an existing standard
protocol.
For packet filtering, the ACL model in RFC 8519 [RFC8519] can
represent supported match and action rules. It does not by itself
express every stateful or application-level requirement. The SCF
Agent checks the target capabilities and traffic binding before
treating a policy realization as equivalent.
3.5. Observation, Analytics, and Policy Memory
Rule status, counters, relevant security events, and path-binding
evidence support outcome verification. Collection and correlation
can follow the Network Telemetry Framework [RFC9232] without
requiring a particular telemetry transport. The observations are
associated with their objective revision, operation, protected scope,
enforcement endpoint, and time.
Ahn, et al. Expires 6 May 2027 [Page 7]
Internet-Draft Agentic AI Security Framework November 2026
The SAF Agent retains objective history, while the SCF Agent retains
target realizations and verification evidence. The distinction
between intended configuration and operational state follows NMDA
[RFC8342] where supported. Configuration readback alone does not
establish live traffic protection, and historical evidence cannot
prove current enforcement.
Monitoring and analysis can be separate supporting services. This
deployment choice does not create another independent policy owner.
Shared records allow the agents to correlate a plan with its results
while keeping desired state, observed state, and inferred conclusions
distinguishable.
4. Agentic Security Management Procedures
Initial policy generation and continuity across network changes form
one lifecycle. The agents retain the objective and policy history
while adapting its low-level realization. The following sequence
describes logical coordination rather than a new routing or device-
management protocol.
4.1. Policy Generation, Delivery, and Continuous Verification
The administrator submits intent to the SAF Agent, which derives the
high-level policy and confirms its scope. The SCF Agent then selects
the current enforcement point, validates a low-level realization, and
coordinates delivery. Figure 2 summarizes the Observe-Reason-Act-
Verify loop used for both initial policy generation and later
adaptation.
Administrator -- intent --- SAF Agent -- policy (A) --- SCF Agent
^ |
| | policy (C)
reported outcome v
| Distributed NSFs
| |
+---- SCF verification -----+
^ telemetry (C)
|
Context / controller -- adapter (B) -- authorized context
Operator inventory ------------------+ endpoint/capability mapping
Figure 2: Agentic Security Control Loop
1. Observe: Collect the intent, current policy, relevant topology
and traffic context, enforcement bindings, and NSF capabilities.
Check that the observations are authorized and sufficiently
current for the decision.
Ahn, et al. Expires 6 May 2027 [Page 8]
Internet-Draft Agentic AI Security Framework November 2026
2. Reason: The SAF Agent establishes or confirms the objective. The
SCF Agent selects an authorized realization and checks
capability, scope, and policy-conflict constraints. The plan
includes the expected outcome and how to verify it.
3. Act: The SCF Agent invokes an authorized executor to install and
activate the approved low-level policy. Responses are correlated
with the intended policy revision, operation, and traffic
binding.
4. Verify: The SCF Agent compares fresh enforcement evidence with
the required outcome and returns the result to the SAF Agent.
Failed or insufficient evidence leads to bounded replanning, an
approved recovery action, or human escalation.
The management exchange remains separate from the actual traffic
path. Initial verification records the active policy, protected
scope, and evidence that traffic uses the intended enforcement point.
An installation response demonstrates configuration progress, but
does not alone establish that the security objective has been
achieved.
For example, an administrator can request that a defined flow remain
blocked while its workload moves between network segments. The SAF
Agent maintains this persistent outcome.
The SCF Agent determines the supported match and action, installs the
current realization, and checks its operation. When the workload or
path changes, the next procedure adapts that realization without
treating the event as an unrelated policy request.
4.2. Policy Continuity during an Enforcement-Point Transition
This procedure covers preparation, policy migration, and verification
when the traffic path or enforcement location changes. It applies to
workload relocation, failover, rerouting, and changes in network
attachment. The source and target can be different devices, service
edges, or virtual security functions; a transition may be planned or
unexpected.
The procedure assumes that the agent can obtain permitted path and
endpoint information and can manage the relevant NSFs. Advance
preparation also requires notice before the actual traffic change.
These are deployment conditions, not capabilities supplied
automatically by the agent architecture. If the enforcement point
does not change, the SCF Agent verifies or updates the existing
binding.
Ahn, et al. Expires 6 May 2027 [Page 9]
Internet-Draft Agentic AI Security Framework November 2026
Observe: The context service reports the relevant change. The SCF
Agent associates it with the protected traffic and current policy
revision and uses inventory to resolve candidate NSF endpoints.
The authorized network controller remains responsible for the
traffic-path decision. The agent checks whether the current
enforcement point remains usable or a target must be prepared.
A location label or a candidate node is not sufficient to determine
the complete enforcement binding. The SCF Agent needs the applicable
traffic scope and a supported mechanism for placing that traffic
through the target NSF. It records whether the mapping is tentative
or selected and refreshes it when the controller chooses another
path.
Context API SAF Agent SCF Agent S-NSF T-NSF
| | | | |
| Authorized | | | |
| path context | | | |
|----------------| context (A) | | |
| |---------------| | |
| | | [Policy repository / |
| | | endpoint inventory] |
| | | Applied-state readback* |
| | |---------------| |
| | | [Validate target plan] |
| | | Install / bind / activate|
| | |--------------------------|
| | |--------------------------|
| | Readiness (A) | revision / binding / result
| |---------------| | |
| Decision (B) | | | |
|----------------| | | |
| [Optional supported controller coordination (B)] |
Interface B: authorized context and controller coordination.
Interface A: proposed agent coordination, not a domain API.
* Source readback is used when available.
Installation and activation are not post-switch verification.
Figure 3: Proposed Policy Preparation and Target NSF Activation
Ahn, et al. Expires 6 May 2027 [Page 10]
Internet-Draft Agentic AI Security Framework November 2026
Reason: The SCF Agent retrieves the desired policy and, when the
source NSF is reachable, compares it with source readback. If
readback is unavailable, it uses the stored desired policy and
available current context, recording uncertainty about the applied
source state. It checks target capabilities, traffic-binding
semantics, and the constraints retained by the SAF Agent. It can
reuse an existing realization, translate it for another target, or
select an authorized alternative.
Act: The SCF Agent installs and activates the applicable low-level
policy and obtains a result for the specified revision and binding.
It reports target readiness to the SAF Agent. If the path controller
supports coordinated preparation, this result may inform the
controller's decision to proceed. The proposed Policy-Ready
Indication does not itself provide routing control or require every
controller to wait.
Readiness is scoped to the prepared destination. If the target
changes, the SCF Agent reevaluates the plan and prepares the selected
enforcement point before claiming protection there. Rules can be
installed in advance only when doing so is safe for the existing
traffic and the deployment can distinguish the applicable bindings.
Verify: The authorized controller performs or confirms the path
change. The target NSF applies the prepared policy to the traffic
using its binding.
The SCF Agent evaluates fresh path and enforcement evidence for the
protected scope and policy revision, then reports the outcome to the
SAF Agent. Missing evidence remains an unknown result rather than a
successful verification.
The source policy binding is retired only after target enforcement is
verified and the old protected path is no longer needed.
During forwarding overlap or concurrent paths, both enforcement
points may need protection. Cleanup must not remove a shared rule
still required by other traffic. Keeping an unused source rule does
not protect packets already taking an unprotected target path.
When a change occurs without advance notice, protection cannot be
guaranteed merely by applying this sequence afterward. The
configured continuity policy must specify whether to retain a
reachable protected path, use a pre-provisioned fallback, block the
affected traffic, or escalate. The agent reports any unresolved
protection state and availability impact.
Ahn, et al. Expires 6 May 2027 [Page 11]
Internet-Draft Agentic AI Security Framework November 2026
Policy migration covers rules and bindings. Connection-tracking,
NAT, or cryptographic state may need additional transfer or recovery
support. A target lacking the required semantics cannot be treated
as equivalent just because it accepts a similar rule. The chosen
continuity action should reflect these capability limits and the
administrator's constraints.
4.3. Policy State and Failure Handling
Accepted objectives, prepared policies, and verified enforcement are
separate results. The SAF Agent tracks the objective, while the SCF
Agent records the desired revision, target realization, action
result, and evidence. A target can be ready before it carries
traffic; the objective is reported as satisfied only when the
relevant evidence supports that outcome.
Correlate migration messages with the objective revision, operation,
and intended traffic binding. A new objective or another path change
may supersede an earlier plan.
Repeated requests should have consistent effects, stale responses
must not advance the current operation, and retry budgets should
prevent repeated oscillation. Failed or uncertain outcomes remain
visible until verification or an authorized recovery resolves them.
4.4. Handling an Unplanned Path Change
An unplanned device failure differs from a coordinated workload move:
the network may already be carrying traffic on another path before
the agent receives the event.
The SCF Agent first determines the active path and whether an
existing or pre-provisioned NSF protects it. The target prepared for
an earlier plan cannot be assumed to match the new destination.
The continuity policy determines the immediate permitted behavior. A
deployment may already have equivalent controls on a backup path or
may support restricting affected traffic until verification
completes.
The agent can select only mechanisms that the authorized controllers
and enforcement points actually provide. An unavailable source
device cannot preserve protection merely because its last reported
policy was correct.
Ahn, et al. Expires 6 May 2027 [Page 12]
Internet-Draft Agentic AI Security Framework November 2026
Recovery establishes a valid binding and verifies the required
outcome on the active path. The SAF Agent reports any period of
unknown or degraded protection, while the SCF Agent records the
failure, action, and evidence.
This separation avoids confusing rapid connectivity recovery with
security-policy continuity. Evaluation should report both service
interruption and any traffic delivered without the intended control.
4.5. Application Example: Mobility-Driven Enforcement Transition
The same procedure can support a mobility event in a 5G deployment.
The following simplified N2 handover diagrams retain the preparation
and execution sequence used in the earlier framework. They
illustrate one application of Section 4.2; they do not restrict the
architecture to mobile networks. The underlying handover procedure
is specified in clauses 4.9.1.3.2 and 4.9.1.3.3 of [TS-23.502].
UE denotes User Equipment; S- and T- identify source and target; NG-
RAN denotes the radio access network. AMF and SMF are the Access and
Mobility Management Function and Session Management Function, and UPF
is the User Plane Function. These domain-specific roles remain
separate from the SAF and SCF Agents.
S-NG-RAN AMF SMF T-UPF T-NG-RAN
| | | | |
| Handover Decision | | |
| Handover Required | | |
|-----------| | | |
| | | | |
| | PDU Session Update Request |
| |------------| | |
| | | Target User-Plane |
| | | Selection and Preparation
| | |----------| |
| | |----------| |
| | | | |
| | PDU Session Update Response |
| |------------| | |
| | Handover Request | |
| |-----------------------------------|
| | Handover Request Acknowledge |
| |-----------------------------------|
Figure 4: N2 Handover Preparation: Illustrative Target-Path Context
Ahn, et al. Expires 6 May 2027 [Page 13]
Internet-Draft Agentic AI Security Framework November 2026
The SMF determines the applicable user-plane arrangement. A handover
does not necessarily select a different UPF or NSF; the target-UPF
exchange shown is conditional on the applicable deployment and
procedure. If the active enforcement point remains unchanged, the
SCF Agent checks the existing policy and binding rather than
migrating it unnecessarily.
Where authorized context identifies a changed enforcement point, the
agents apply Figure 3. A 5G Application Function adapter can provide
permitted context through NEF or another authorized exposure API,
while operator inventory resolves the corresponding NSF endpoints.
NEF exposure is not the SAF-SCF interface, and an exposed service or
location identifier is not by itself an NSF management address.
Target preparation uses the desired policy and available source
readback, capability validation, and a readiness result. Readiness
does not introduce a new N2 message or a mandatory handover barrier.
Any preparation acknowledgment affects the network transition only
where an existing, authorized integration supports that coordination;
otherwise, the configured continuity policy governs the interval.
UE S-RAN T-RAN AMF SMF T-UPF
| | | | | |
| | Handover Command | | |
| |---------------------| | |
| Handover Command | | | |
|---------| | | | |
| | Uplink RAN Status Transfer | |
| |---------------------| | |
| | | Downlink RAN Status Transfer |
| | |----------| | |
| [Target-cell synchronization] | | |
| Handover Confirm | | | |
|--------------------| | | |
| | | Handover Notify | |
| | |----------| | |
| | | | Session context update
| | | |----------| |
| | | | | N4 update|
| | | | |----------|
| | | [Updated user-plane path] |
Figure 5: N2 Handover Execution: Illustrative User-Plane Path Update
Ahn, et al. Expires 6 May 2027 [Page 14]
Internet-Draft Agentic AI Security Framework November 2026
After the network updates the user-plane path, the SCF Agent verifies
the selected traffic binding and live enforcement at the applicable
NSF. Handover completion alone is not security verification. Old
bindings are retired only after verification and confirmation that
they are no longer needed. A UPF path update and migration of a
stateful security function are separate operations.
5. Multi-Domain Coordination
The SAF Agent can maintain a shared security objective across
different network environments. Figure 6 illustrates a 5G domain
alongside a cloud or campus domain. Each has a domain-local SCF
Agent and its own enforcement functions. The example shows how the
general framework can be applied to different networks while
preserving a common objective.
+-----------------------------+
| SAF Agent: shared objective |
+--------------+--------------+
|
+----------+-----------+
| |
Agent interface A Authorized domain interface
| |
+-------v------+ +---------v----------------+
| 5G SCF Agent | | Cloud / campus SCF Agent |
+-------+------+ +---------+----------------+
| |
+-------v------+ +---------v----------------+
| UPF / NSFs | | Cloud / campus NSFs |
+--------------+ +--------------------------+
5G context: SAF AF adapter --- NEF / authorized 5G exposure.
All domains return local outcome evidence to the SAF Agent.
Figure 6: Shared Objective with Domain-Local Security Control
In the 5G example, a User Plane Function (UPF) is associated with the
applicable NSFs. Authorized context can be obtained through an
Application Function (AF) integration and the Network Exposure
Function (NEF), consistent with the 5G architecture [TS-23.501]. The
NEF provides 5G capability exposure; it is not a general transport
for SAF-SCF objective messages. Cloud and campus domains use their
own authorized interfaces.
The shared objective does not grant access across administrative
boundaries. Each SCF Agent validates local permissions and
capabilities and reports whether its realization is satisfied,
Ahn, et al. Expires 6 May 2027 [Page 15]
Internet-Draft Agentic AI Security Framework November 2026
unsupported, or unverifiable. The SAF Agent combines these outcomes
and reports partial protection when necessary. Success in one domain
does not establish an end-to-end result.
The architecture does not assume an atomic transaction across
independently managed domains. Coordinated recovery retains local
authority and the shared constraints. A failure in one domain cannot
be rolled back by changing an unrelated domain's rule; each local
outcome must remain visible to the SAF Agent.
6. Security Considerations
Authenticate administrators, agents, context providers, and
enforcement functions, and authorize actions for the relevant
objective and traffic scope. Protect policy, context, and telemetry
exchanges against disclosure, modification, and replay. On supported
NETCONF or RESTCONF endpoints, NACM [RFC8341] can restrict management
operations and data access. It does not define SAF-SCF
authorization; other interfaces need explicit authorization
mechanisms.
Treat intent text, telemetry, and retrieved records as potentially
untrusted data. AI-generated interpretations and plans must pass
independent scope, capability, and policy checks before an executor
uses credentials or changes enforcement. A model cannot expand its
own authority by generating instructions or interpreting a lower-
trust observation as an operator command.
Bind readiness and evidence to the current objective revision,
operation, target, and traffic scope. Keep audit records of the
observations, decisions, actions, and outcomes while minimizing
unnecessary user or workload identifiers.
Missing or stale telemetry is not proof of protection. Source
cleanup requires both target verification and confirmation that the
old binding is no longer required.
Define fallback, rollback, and human-escalation behavior in advance.
Routing ownership and security-rule ownership remain explicit, and
availability pressure alone does not authorize a weaker security
objective.
If no permitted plan can satisfy the request, report the limitation
and use the approved continuity action. An agent cannot undo a
completed network transition simply by restoring an obsolete rule.
Ahn, et al. Expires 6 May 2027 [Page 16]
Internet-Draft Agentic AI Security Framework November 2026
7. Conclusion
This document develops distributed security management into an
Agentic AI framework for heterogeneous networks. The SAF Agent
maintains the objective; the SCF Agent plans, executes, and verifies
its realization at distributed NSFs.
Policy generation and continuity across path or enforcement-point
changes form one Observe-Reason-Act-Verify lifecycle. Existing
management, modeling, telemetry, and access-control mechanisms
provide integration foundations while agent interoperability remains
further work.
8. IANA Considerations
This document has no IANA actions.
9. References
9.1. Normative References
[RFC9315] Clemm, A., Ciavaglia, L., Granville, L. Z., and J.
Tantsura, "Intent-Based Networking - Concepts and
Definitions", RFC 9315, DOI 10.17487/RFC9315, October
2022, <https://www.rfc-editor.org/info/rfc9315>.
9.2. Informative References
[RFC6241] Enns, R., Ed., Bjorklund, M., Ed., Schoenwaelder, J., Ed.,
and A. Bierman, Ed., "Network Configuration Protocol
(NETCONF)", RFC 6241, DOI 10.17487/RFC6241, June 2011,
<https://www.rfc-editor.org/info/rfc6241>.
[RFC7950] Bjorklund, M., Ed., "The YANG 1.1 Data Modeling Language",
RFC 7950, DOI 10.17487/RFC7950, August 2016,
<https://www.rfc-editor.org/info/rfc7950>.
[RFC8040] Bierman, A., Bjorklund, M., and K. Watsen, "RESTCONF
Protocol", RFC 8040, DOI 10.17487/RFC8040, January 2017,
<https://www.rfc-editor.org/info/rfc8040>.
[RFC8341] Bierman, A. and M. Bjorklund, "Network Configuration
Access Control Model", RFC 8341, DOI 10.17487/RFC8341,
STD 91, March 2018,
<https://www.rfc-editor.org/info/rfc8341>.
Ahn, et al. Expires 6 May 2027 [Page 17]
Internet-Draft Agentic AI Security Framework November 2026
[RFC8342] Bjorklund, M., Schoenwaelder, J., Shafer, P., Watsen, K.,
and R. Wilton, "Network Management Datastore Architecture
(NMDA)", RFC 8342, DOI 10.17487/RFC8342, March 2018,
<https://www.rfc-editor.org/info/rfc8342>.
[RFC8345] Clemm, A., Medved, J., Varga, R., Bahadur, N.,
Ananthakrishnan, H., and X. Liu, "A YANG Data Model for
Network Topologies", RFC 8345, DOI 10.17487/RFC8345, March
2018, <https://www.rfc-editor.org/info/rfc8345>.
[RFC8519] Jethanandani, M., Agarwal, S., Huang, L., and D. Blair,
"YANG Data Model for Network Access Control Lists (ACLs)",
RFC 8519, DOI 10.17487/RFC8519, March 2019,
<https://www.rfc-editor.org/info/rfc8519>.
[RFC9232] Song, H., Qin, F., Martinez-Julia, P., Ciavaglia, L., and
A. Wang, "Network Telemetry Framework", RFC 9232,
DOI 10.17487/RFC9232, May 2022,
<https://www.rfc-editor.org/info/rfc9232>.
[TS-23.501]
3GPP, "System architecture for the 5G System (5GS)", 3GPP
TS 23.501 V18.5.0, May 2024,
<https://www.etsi.org/deliver/
etsi_ts/123500_123599/123501/18.05.00_60/
ts_123501v180500p.pdf>.
[TS-23.502]
3GPP, "Procedures for the 5G System (5GS)", 3GPP TS 23.502
V18.5.0, May 2024, <https://www.etsi.org/deliver/
etsi_ts/123500_123599/123502/18.05.00_60/
ts_123502v180500p.pdf>.
Appendix A. Changes from draft-ahn-nmrg-5g-security-i2nsf-framework-02
This document revises draft-ahn-nmrg-5g-security-i2nsf-framework-02.
The major changes are as follows:
* This version introduced Agentic AI for intent interpretation,
policy planning, execution, and continuous verification.
* It broadened the scope to mobile and wireless networks including
5G networks and WiFi networks. It reduced dependence on the I2NSF
framework, while retaining 5G use case as an example.
* It refined policy-continuity procedures, including target
readiness, enforcement verification, and failure handling.
* It updated security considerations, integration with existing
standards, and references.
Ahn, et al. Expires 6 May 2027 [Page 18]
Internet-Draft Agentic AI Security Framework November 2026
Acknowledgments
This work was supported by Institute of Information & Communications
Technology Planning & Evaluation (IITP) grant funded by the Korea
Ministry of Science and ICT (MSIT) (No. RS-2024-00398199).
Contributors
This document has benefited from discussions in NMRG and from inputs
and texts by Linda Dunbar (Futurewei), Yong-Geun Hong (Daejeon
University), and Joo-Sang Youn (Dong-Eui University). The authors
sincerely appreciate their contributions.
The following is the coauthor of this document:
Mose Gu
Department of Computer Science & Engineering
Sungkyunkwan University
2066 Seobu-Ro, Jangan-Gu
Suwon, Gyeonggi-Do 16419
Republic of Korea
Phone: +82 31 299 4106
Email: rna0415@skku.edu
URI: http://iotlab.skku.edu/people-Moses-Gu.php
Authors' Addresses
Yoseop Ahn (editor)
Department of Computer Science & Engineering
Sungkyunkwan University
2066 Seobu-Ro, Jangan-Gu
Suwon, Gyeonggi-Do 16419
Republic of Korea
Phone: +82 31 299 4106
Email: ahnjs124@skku.edu
URI: http://iotlab.skku.edu/people-Ahn-Yoseop.php
Jaehoon Paul Jeong (editor)
Department of Computer Science & Engineering
Sungkyunkwan University
2066 Seobu-Ro, Jangan-Gu
Suwon, Gyeonggi-Do 16419
Republic of Korea
Phone: +82 31 299 4957
Email: pauljeong@skku.edu
URI: http://iotlab.skku.edu/people-jaehoon-jeong.php
Ahn, et al. Expires 6 May 2027 [Page 19]
Internet-Draft Agentic AI Security Framework November 2026
Younghan Kim
School of Electronic Engineering
Soongsil University
369, Sangdo-ro, Dongjak-gu
Seoul 06978
Republic of Korea
Email: younghak@ssu.ac.kr
Ahn, et al. Expires 6 May 2027 [Page 20]