Match each cloud managed guardrail product to its provider

Match pairs·Medium·4.0 · 0·~2 min·Asked atBrowserbase Comet Ml Siemens

Attempt it

Drag each answer to line up with its matching prompt

Prompt Shields

Google Cloud

Bedrock Guardrails

Amazon AWS

Model Armor

Microsoft Azure

NeMo Guardrails

Meta

Llama Guard 4 / Prompt Guard 2

OpenAI

Moderation API and safety filters

NVIDIA

Each major cloud and model vendor has its own guardrail product, and the names all sound similar. The three hyperscaler cloud-managed shields are **Azure Prompt Shields** from Microsoft, **AWS Bedrock Guardrails** from Amazon, and **Google Model Armor** from Google Cloud. All three sit between your application and the hosted model API. They offer input filtering, output filtering, and policy categories for things like harmful content, PII, and prompt injection. The open-source and vendor-branded side has **NVIDIA NeMo Guardrails**, an orchestration framework using a small DSL called Colang. Meta ships **Llama Guard 4** for content safety and **Prompt Guard 2** for injection detection, both as open-weights models you self-host. **OpenAI Moderation** is a free endpoint on OpenAI's API. Anthropic's API-side safeguards run inside its own API. Neither is sold as a separate product. The match pattern: cloud providers brand with cloud-themed words (Shields, Guardrails, Armor), and model vendors name after their model family (Llama Guard, NeMo).

**Six vendors, six product lines.** - Microsoft Azure: **Prompt Shields** (part of Azure AI Content Safety). Screening for user prompt attacks and document (indirect) attacks, used inline in Azure OpenAI content filtering or through a standalone API. - Amazon AWS: **Bedrock Guardrails**. Six safeguard policies (content filters, denied topics, word filters, sensitive information filters, contextual grounding checks, Automated Reasoning checks), applied to any model invoked through Bedrock and, through the ApplyGuardrail API, to other models. - Google Cloud: **Model Armor**. Screening and policy enforcement for prompts and responses of any model, including models on other clouds. - NVIDIA: **NeMo Guardrails**. Open-source orchestration with Colang as the DSL, used in self-hosted stacks on GPU clusters. - Meta: **Llama Guard 4** (broad safety classifier, multimodal) and **Prompt Guard 2** (specialised injection/jailbreak detector). Open-weights, self-hostable. - OpenAI: **Moderation API**, a free endpoint that classifies text and images against OpenAI's harm categories. **In practice.** Security and procurement reviews often ask which managed shield a deployment uses. The brand names map directly to compliance posture. Bedrock Guardrails on Bedrock-hosted Claude means a single-vendor audit trail. NeMo plus Llama Guard self-hosted on the customer's GPU means data residency in the customer's boundary. The right choice depends on the deployment, not on which product is 'best.'

TL;DR

Azure ships Prompt Shields, AWS ships Bedrock Guardrails, Google ships Model Armor, NVIDIA ships NeMo Guardrails, Meta ships Llama Guard plus Prompt Guard, OpenAI ships Moderation.

Memory aid

Easy to grasp

Imagine three big grocery chains all selling their own brand of bleach. Same job, different bottles, slightly different smells. Cloud guardrails are like that. Each big cloud has rebottled the safety layer with its own brand name so they can sell it next to their model hosting. The OSS world has its own brands too. Meta makes the safety-classifier bottles and NVIDIA makes the orchestration bottle. OpenAI runs its own first-party detector behind the API. The job each of them does is the same shape: filter inputs, filter outputs, and refuse policy-violating prompts. The price tags, the latency, and the supported policy categories are what differ.

Key concepts

Safety & Guardrails

Concept explanation~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

The cloud-managed guardrail landscape is small enough to memorise and confusing enough that the products are often cross-attributed. Knowing the names and the vendor behind each one is the minimum. The axes along which they differ (data residency, customisation, latency, failure ownership) matter more when building a production safety layer.

Cloud providers brand their guardrails with cloud-flavoured words (Shields, Guardrails, Armor). Model vendors brand theirs with their model family (NeMo, Llama Guard, Prompt Guard). The pattern holds across every major name.

The six vendor product lines

The hyperscaler trio

The three big cloud providers each ship a managed guardrail product:

Azure Prompt Shields is Microsoft's offering, part of the broader Azure AI Content Safety service. It screens for user prompt attacks and document (indirect) attacks and integrates with Azure OpenAI deployments. The naming hints at scope: 'Shields' suggests the inline protection role.
AWS Bedrock Guardrails is Amazon's product, layered around any model invoked through Bedrock, and through the ApplyGuardrail API around other models too. It exposes six policy types: content filters, denied topics, word filters, sensitive information filters, contextual grounding checks and Automated Reasoning checks. The branding follows AWS service-naming convention.
Google Cloud Model Armor is Google's equivalent. It screens prompts and responses for any model, including models on other clouds. It checks for prompt injection and jailbreak, sensitive data, malicious URLs and harmful content. Naming is again service-specific.

The OSS and vendor-branded set

NVIDIA NeMo Guardrails is the open-source orchestration framework, with the Colang DSL as the rule language. It is not a classifier itself. It calls classifiers (Llama Guard, custom models, regex) as rails.
Meta Llama Guard 4 is the broad safety classifier with a configurable taxonomy and multimodal coverage. Meta Prompt Guard 2 is the specialised injection/jailbreak detector. Both ship as open-weights models you self-host.
OpenAI Moderation is the first-party content classifier exposed at the OpenAI platform API. Anthropic documents no separate moderation endpoint; its classifier safeguards run inside its own API.

A note on smaller players

The vendor landscape also includes Lakera Guard (a hosted screening API, now owned by Check Point), Guardrails AI (open-source Python SDK with a validator hub), and Vijil (evaluation, red teaming and the open-source Vijil Dome guardrails). They sit alongside the big six rather than replacing them.

Cloud-managed versus self-hosted: the real axis

The hybrid stack most enterprises converge on

Common confusions worth catching

Situations where this technique stops working.

2–4 min · Everything important, quickly.

Real products, models, and research that use this idea.

Bedrock Guardrails' ApplyGuardrail API screens prompts and responses for any model, including self-hosted and third-party ones, without invoking a foundation model.
Azure Prompt Shields detects user prompt attacks and document attacks. Its standalone API returns attackDetected and leaves blocking to the application.

What an interviewer would ask next. Try answering before peeking at the approach.

QHow would you choose between Bedrock Guardrails and self-hosted Llama Guard 4 for a B2B SaaS shipping in three cloud regions?

Bedrock Guardrails wins on operational simplicity and audit trail across the three regions if all three are AWS. Self-hosted wins on customisation and on multi-cloud portability. A common pattern is Bedrock for default coverage and self-hosted Prompt Guard 2 as a secondary on routes that handle customer PII.

2 more follow-ups an interviewer would ask next. Sign in to reveal them.

Red flags & common mistakes

The phrases that signal junior thinking. Click to expand.

Most common mistake

Confusing the cloud-managed shields with each other or thinking Llama Guard is Microsoft's product. Llama Guard ships from Meta and is self-hostable, not a cloud-managed offering.

60 second bullets to scan on the way to the call.

Six vendors and their guardrail product lines
Distinction between cloud-managed shields and self-hosted classifiers

Primary sources. Browse if you want the original framing.