Load Balancers, ADCs, WAAP, and AI Security: What Goes Where?

Yuta Matsuyama

Security Researcher, TECHFUND Inc.

Updated:

Who this is for: This article is intended for development and security professionals operating web services or AI features, and is based on publicly available materials as of November 13, 2026. Features that use AI to detect attacks are different from features that protect AI applications. Here, AI security refers to the latter.

The practical answer

“We already have a load balancer. Do we also need WAAP?” “Will an AI firewall protect our APIs too?” These four may look like similar product categories, but the starting points for evaluating them differ. Load balancers focus on distributing traffic; ADCs on controlling delivery; WAAP on detecting and mitigating attacks against websites and APIs; and AI security on addressing risks specific to AI systems. Features overlap depending on the product, so it is important not to decide whether one can replace another based on names alone.

The Four Roles and the Boundaries of Evaluation

This is a classification by function. In practice, one product may fulfill multiple roles, or multiple products may be combined. For example, F5’s ADC materials describe TLS termination, L7 routing, and policy control in addition to traffic distribution. However, check individually which products and contracts include these capabilities.

The Four Roles and the Boundaries of Evaluation
TypeMain roleLimitations to check
Load balancer (LB)Distributes traffic across multiple destinations and selects a forwarding destination based on their statusDoes not necessarily include attack detection or permission checks within applications
ADC (Application Delivery Controller)Combines delivery control, including load balancing, with TLS termination, routing, monitoring, policies, and other capabilitiesIncluded security features and operational scope vary by product and configuration
WAAP (Web Application and API Protection)Combines WAF, API protection, bot mitigation, DDoS mitigation, and other capabilities to reduce threats to publicly exposed surfacesDoes not protect every API, internal communication, or business-level authorization decision
AI securityReduces risks involving model inputs and outputs, retrieved documents, agent tool use, and related areasVisibility varies depending on supported models, connection methods, and inspection targets

Separate the Traffic Entry Point from What Happens Inside the Application

LBs include L4 types that handle connection information such as TCP, and L7 types that use information such as HTTP paths. For example, AWS’s architecture example allows forwarding from a Network Load Balancer to an Application Load Balancer, combining this with per-request distribution at L7. Conditions include placing them in the same VPC and the same account. This does not mean that you must always deploy multiple LBs. This example does not mean that an LB can only distribute traffic. AWS ALB also supports user authentication, but whether a user can access a specific order is a separate authorization decision.

On the WAAP side, for example, Cloudflare’s documentation describes how WAF managed rules, rate limiting, bot mitigation, and L7 DDoS mitigation work together. What matters is not only whether a feature exists, but whether traffic to the relevant hosts and APIs actually passes through it, what inspection is skipped by exception settings, and how false positives are tuned. The same documentation also states that WAF custom and managed rules do not apply to certain subrequests from Workers and similar sources within the same zone.

The following is one example of an architecture. The actual order and degree of integration vary by service.

Separate the Traffic Entry Point from What Happens Inside the Application
  1. Users
  2. Protection for publicly exposed surfaces, such as WAAP
  3. Forwarding and delivery control by a load balancer / ADC
  4. Authentication and permission checks for applications and APIs
  5. AI features (retrieved documents → model → tool calls)

Even with Overlapping Defenses, Application Permission Checks Remain Necessary

For example, consider an API through which a logged-in user accesses /orders/123. If the user can change the order ID to another user’s order and access it, the request format may still appear normal. Alongside WAAP inspection, the application must determine, for each order, “Is this user allowed to view this order?” OWASP’s API security materials categorize this issue as Broken Object Level Authorization (BOLA).

AI features have a similar boundary. Not only instructions entered directly into a question field, but also instructions embedded in documents retrieved through search, can change a model’s behavior. OWASP’s explanation of prompt injection gives examples of indirect attacks through external documents. In addition to inspecting inputs and outputs, you must keep tool permissions to the minimum necessary and enforce authorization for each protected operation within the application or through a trusted authorization infrastructure. For high-impact operations, also consider requiring additional approval. Do not assume that deploying an inspection product can prevent these attacks completely.

From the perspective of integrating code analysis into development, what matters is not only whether input was blocked at the entry point, but also which application operations and data the input that passed through can reach. Cross-check the network configuration against the authorization points in the code, and identify processing that is not visible from entry-point checks.

Four Things to Check Before Choosing

1. Which paths will be protected? List which of the following are actually inspected: public websites, APIs, internal communications, communications with AI models, retrieved documents, and tool calls.

2. What can it assess? Check whether it can assess not only traffic volumes and patterns, but also the relationships among users, tenants, and the data being accessed. If the information needed to make that decision is unavailable, application-side authorization is necessary.

3. What happens during failures or false positives? For each product, check whether traffic is blocked or allowed through when the inspection service is unavailable, along with logging, exception settings, and recovery procedures.

4. Who owns operational responsibility? Decide who will review delivery settings, detection rules, application permissions, and AI tool permissions, respectively.

A good approach is to first mark these boundaries on your existing architecture diagram and then select solutions in the order needed to fill the gaps. Situations requiring an LB are different from those requiring a review of the application’s authorization implementation. When adding WAAP or AI-oriented protection, comparing options only after checking existing capabilities and their visibility also makes overlaps and gaps easier to identify.

Common questions

Does a load balancer make WAAP unnecessary?

A load balancer primarily distributes traffic. WAAP inspection capabilities vary by product, so verify the scope actually protected.

Can AI inspection completely prevent prompt injection?

Do not assume complete prevention. Also check how retrieved documents are handled, tool permissions, and authorization for each operation.

How Hi AUDIT Agent can help

In addition to protecting publicly exposed surfaces, another approach is to review code during development. Hi AUDIT Agent integrates with IDEs and helps examine data flows from inputs to processing destinations in web application code, as well as information on known vulnerabilities. It does not replace WAAP protection for live traffic or the work of defining application-specific authorization rules, so use it with these roles clearly separated.

Create a Hi AUDIT Agent account