🛡️ Azure AI + Machine Learning • DCSA Customer Briefing

Azure AI
Content Safety

The AI safety layer that detects harmful user-generated and AI-generated content — across text and images — so you can ship generative AI with confidence.

Microsoft DCSA Digital Cloud Solution Architect
Customer Briefing • Grounded in Microsoft Learn documentation

→ or Space to begin · Notes button, bottom-left · F fullscreen

Open strong. This is the safety layer for GenAI — one service, text + image, harmful-content detection plus prompt-attack protection. Frame the whole talk as "how do you deploy AI responsibly and stay compliant." Set expectation: ~15 min + live demo. Don't read the slide — say the value prop and move.

Azure AI Content Safety

The Problem

Generative AI created a new attack surface

Every app that accepts user input or emits AI-generated output can now surface hate, sexual, violent, or self-harm content — or be manipulated into it. Manual moderation and hand-built models don't scale, and regulators expect controls.

Harmful content, both directions. User prompts and model completions can contain offensive, risky, or undesirable material.
Prompt attacks & jailbreaks. Users deliberately exploit system vulnerabilities to make an LLM break its rules.
Trust & compliance. Content filtering helps apps comply with regulations and keep the intended environment for users.

"Content filtering software can help your app comply with regulations or maintain the intended environment for your users."— Azure AI Content Safety documentation

0

Harm categories: Hate, Sexual, Violence, Self-Harm

0–0

Severity scale for granular, risk-based filtering

2

Directions guarded — inbound prompts & outbound AI output

1

Unified safety service across your whole AI stack

DCSA Customer Briefing

Anchor in the customer's reality: the moment they added a chatbot or accept UGC, they own a moderation problem in both directions. Use the doc's exact framing — compliance + intended environment. Don't fearmonger; position Content Safety as the pragmatic control. Maps to Key Insights #1 and #2.

Azure AI Content Safety

What is it

A cloud AI service that detects harmful content

Azure AI Content Safety is an AI service that detects harmful user-generated and AI-generated content in applications and services. It provides text and image APIs to detect material that's harmful, plus an interactive Content Safety Studio to explore and try sample code.

Where it fits — a dedicated safety & governance layer inside the Azure AI + Machine Learning portfolio, sitting between your users and your models:

👥

Users & UGC

→

🛡️

Azure AI Content Safety

→

🧠

LLM / GenAI App

→

✅

Safe Experience

To use it you need an Azure subscription and a Content Safety resource created in a supported region. Note: the service can't be used to detect illegal child exploitation images.

DCSA Customer Briefing

Keep the definition verbatim-close to the doc — credibility matters. Stress "text AND image" and that Studio lets non-developers explore before writing code. The "where it fits" flow is the mental model to repeat all deck: it's the layer between users and your model. Maps to Key Insights #1, #9, #11.

Azure AI Content Safety

Core Capabilities

One service, six ways to keep AI safe

📝

Analyze Text

Scans text for Hate, Sexual, Violence & Self-Harm with multi-severity levels (0–7).

GA

🖼️

Analyze Image

Scans images for the same four harm categories with multi-severity risk levels (0/2/4/6).

GA

🛑

Prompt Shields

Detects jailbreak (user prompt attacks) and indirect / document injection attacks on your LLM.

GA

🎯

Groundedness Detection

Detects when an LLM answer drifts from your sources — and can auto-correct the ungrounded text back to fact.

Preview

©️

Protected Material

Scans AI output for known text (lyrics, articles, recipes) and code from public repositories.

Text GA Code Preview

🧩

Custom Categories

Create & train your own categories — standard for trained, rapid for emerging patterns.

Preview

Plus built-in & custom blocklists, Task Adherence for AI agents (Preview), and Multimodal image-with-text analysis (Preview).

DCSA Customer Briefing

Don't list six flatly. Land three buckets: (1) content moderation — text + image; (2) GenAI protection — Prompt Shields, Groundedness, Protected Material; (3) customization — custom categories + blocklists. Call out GA vs Preview honestly. Maps to Key Insights #2, #3, #4, #5, #6, #12.

Azure AI Content Safety

How It Works · Live Demo

The safety layer between your users and your model

1

Ingest

User prompt or content enters your app.

2

Send

App calls the Content Safety API.

3

Analyze

Models score 4 harm categories 0–7 & detect attacks.

4

Decide

Compare severity to your threshold + blocklists.

5

Act

Allow safe content, block/escalate harmful.

🧪 Try it: severity threshold playground

Front-end visualization of the doc's severity logic (0/2/4/6). Not a real API call — it shows how a threshold decides Allowed vs Blocked per category.

Block at severity ≥4 · Medium

0 Safe2 Low4 Medium6 High

📊 Category & severity result

Four harm categories, each with a simulated severity, evaluated against your threshold.

—

DCSA Customer Briefing

Trace the animated flow left→right: user → app → Content Safety → decision. Then run the demo: type "threat + hate" and drag the slider 6→2, watch categories flip to Blocked. Teach the doc's rule: text supports full 0–7, results can be trimmed to 0/2/4/6; recommended starting threshold is severity 4 for new projects. Reiterate the demo is illustrative — the real service returns severity per category and YOU enforce the threshold. Maps to Key Insights #2, #7.

Azure AI Content Safety

Setup in the Azure Portal

From zero to endpoint in five steps

1

Have an Azure subscription

You need an active Azure subscription — you can create one for free.

2

Create a Content Safety resource

In the Azure portal, create a Content Safety resource. Enter a unique name, then select your subscription.

3

Choose resource group, region & tier

Select a resource group, a supported region, and a supported pricing tier (F0 free or S0 standard), then select Create.

4

Grab your key & endpoint

Once deployed, copy the resource key and endpoint — these authenticate your API calls. Managed Identity is enabled automatically.

5

Explore in Content Safety Studio

Open Studio to moderate text/image, tune sensitivity, manage blocklists, then View Code / export and deploy into your app.

DCSA Customer Briefing

This is the "it's easy" moment. Five steps, all from the doc. Emphasize F0 exists for free PoCs. Highlight Managed Identity is on automatically — a security win they get for free. Studio's "View Code / export" bridges click-ops to production code. Maps to Key Insights #7, #8, #9.

Azure AI Content Safety

Integration

A few lines to embed the safety layer

analyze_text.py

▸ SAMPLE RESPONSE (categoriesAnalysis)

{
  "categoriesAnalysis": [
    { "category": "Hate", "severity": 2 },
    { "category": "SelfHarm", "severity": 0 },
    { "category": "Sexual", "severity": 0 },
    { "category": "Violence", "severity": 6 }
  ]
}

Meets developers where they are

GA client SDKs for C#, Python, Java & JavaScript — plus REST, Azure PowerShell & CLI.

Request options include categories, blocklistNames, haltOnBlocklistHit and outputType (Four/EightSeverityLevels).

DCSA Customer Briefing

Toggle the tabs live and hit Copy. Story: create client with endpoint + key (or Entra ID), one call, iterate categoriesAnalysis, enforce your threshold. outputType picks 4 vs 8 severity granularity; blocklists add domain terms. "This is an afternoon, not a project." Maps to Key Insights #2, #6.

Azure AI Content Safety

Real-World Scenario · 1 of 3

Social & UGC moderation at scale

💬

Social / Messaging Platform

Customer profile

A social messaging platform moderating images and text added by millions of users in real time.

Benefit

Automatically flag hate, sexual, violent & self-harm content at scale — tune severity per community and layer in custom blocklists for profanity and trends.

Features used

Analyze TextAnalyze ImageBlocklistsSeverity tuning

DCSA Customer Briefing

The doc explicitly names social messaging platforms that moderate user images/text. Emphasize both modalities + custom blocklists + per-community severity. If the customer is consumer/UGC, go deep here. Maps to Key Insights #1, #2, #6.

Azure AI Content Safety

Real-World Scenario · 2 of 3

GenAI customer-service chatbot filtering

🤖

Enterprise Support Assistant

Customer profile

An enterprise filtering both user prompts and AI-generated responses in a customer-service assistant.

Benefit

Prompt Shields stops jailbreaks and indirect injection hidden in retrieved documents; Groundedness flags — and can auto-correct — answers that drift from your sources; Protected Material keeps copyrighted text & code from leaking.

Features used

Prompt ShieldsGroundednessProtected MaterialAnalyze Text

DCSA Customer Briefing

This is the highest-value GenAI story — the "safe LLM app" trifecta: Prompt Shields + Groundedness + Protected Material. Perfect for any customer building a copilot/RAG assistant. Maps to Key Insights #3, #4, #5.

Azure AI Content Safety

Real-World Scenario · 3 of 3

In-game chat & artifact safety

🎮

Gaming Platform

Customer profile

A gaming company moderating user-generated game artifacts and chat rooms in real time.

Benefit

Detect misconduct & prevent inappropriate behavior in chats and forums; custom categories (rapid) catch emerging harmful patterns fast.

Features used

Analyze TextCustom Categories (rapid)BlocklistsReal-time

DCSA Customer Briefing

The doc names gaming companies moderating game artifacts + chat rooms. Rapid custom categories are the differentiator — define an emerging pattern and scan immediately. Maps to Key Insights #6.

Azure AI Content Safety

Pricing & Tiers

Start free, scale on standard

F0 · Free

Prototype & validate at no cost

✓ Moderation APIs (text & image): 5 RPS
✓ Prompt Shields: 5 RPS · Protected Material: 5 RPS
✓ Custom categories (rapid) & Multimodal: 5 RPS
✓ Groundedness detection: N/A on F0

S0 · Standard

Production throughput

✓ Moderation, Prompt Shields, Protected Material: 1000 RP10S
✓ Groundedness detection: 50 RPS
✓ Custom categories (standard): 5 RPS · Multimodal: 10 RPS
✓ Need more? Contact Microsoft to request a higher rate.

Input limit (from the doc)	Value	Feature	Value
Analyze text — max length	10K chars	Prompt Shields — max prompt	10K chars
Analyze image — max size	4 MB	Prompt Shields — documents	≤5, 10K chars total
Image dimensions	50×50 – 7200×7200 px	Groundedness — sources / call	55,000 chars

Two tiers: F0 and S0. Exact per-transaction prices — see the Azure pricing page [Verify: pricing page].

DCSA Customer Briefing

The doc gives tiers (F0/S0), rate limits and input limits — NOT dollar figures, so mark price [Verify] and point to the Azure pricing page. Lead with "F0 is free — build your PoC this week." Groundedness is S0-only. Promise a follow-up pricing estimate; never guess numbers. Maps to Key Insights #7, #11.

Azure AI Content Safety

Responsible AI · Security · Compliance

Safe by design, secure by default

🔐

Identity & Access

Microsoft Entra ID (API & SDK)
Managed Identity — on automatically
RBAC: Cognitive Services User + Reader

🗝️

Data Protection

Encryption of data at rest
Customer-managed keys (CMK / BYOK)
Create, rotate, disable & revoke keys

⚖️

Responsible AI

Transparency Note & Microsoft AI Principles
Severity-based, explainable flags
Models kept current with new attacks

🌍

Coverage & Limits

Auto language detection (text & Prompt Shields)
Specially trained: EN, ZH, FR, DE, ES, IT, JA, PT
Can't detect illegal child exploitation imagery

Some models are English-only (Protected Material, Groundedness, Custom categories standard). Microsoft continuously updates models and monitors production traffic; false positives/negatives can be submitted to improve models for your use case.

DCSA Customer Briefing

The trust slide for security/compliance stakeholders. Hit Entra ID + Managed Identity + RBAC + CMK. Reference the Transparency Note for Responsible AI. Be candid about English-only models and the child-exploitation exclusion — candor builds credibility. Maps to Key Insights #8, #10, #11.

Azure AI Content Safety

DCSA Talking Points & Objection Handling

What to say — and how to answer back

SAY
Problem

Name the risk

"The moment you added AI or accept user content, you own a moderation problem in both directions."

SAY
What

Position the layer

"Think of it as a safety layer between your users and your model — text and image, one service."

SAY
Demo

Make it tangible

"You set the threshold; the service returns a severity per category. Start at 4, then tune."

SAY
Pricing

Remove friction

"There's a free F0 tier — prove value this week before spending a dollar."

Q&A
"Keyword filter?"

Objection

"No — ML models rate severity 0–7 across four categories, plus prompt-attack, groundedness & protected-material detection. Blocklists are optional on top."

Q&A
"Our language?"

Objection

"Language is auto-detected; 8 specially-trained languages, others work with varying quality. Some models are English-only — we'll test in the PoC."

Q&A
"Is it GA?"

Objection

"Core text/image, Prompt Shields & Protected Material (text) are GA. Groundedness, custom categories, multimodal are preview — plan accordingly."

Q&A
"Exact price?"

Objection

"Two tiers, F0/S0; per-transaction pricing is on the Azure pricing page — I'll send a tailored estimate."

DCSA Customer Briefing

This is your teleprompter — you can run the whole deck from here. Each SAY line is one breath; each Q&A is a rehearsed rebuttal. Adapt vocabulary to the room (CISO vs dev lead vs product). Never invent pricing or unlisted features.

Azure AI Content Safety

Next Steps

Your path to production

1

Free tier

Spin up a Content Safety resource on F0 — no cost, supported region, minutes to deploy.

2

Explore Studio

Try text & image moderation, tune sensitivity and blocklists — no code required.

3

Guided PoC

We wire Content Safety into one real flow — a chatbot or UGC pipeline — and measure block rate.

4

Scale on S0

Move to standard throughput, add Entra ID + CMK, and roll out across your AI estate.

📘 Read the overview ⚡ Text quickstart

DCSA Customer Briefing

Close with a concrete ask: a joint PoC on F0 wired into one real flow, with block-rate as the success metric. Offer a Studio demo session and a follow-up pricing estimate. Get a date on the calendar before you leave the room. Maps to Key Insights #7, #9.

Azure AI Content Safety

Resources

Learn more — straight from Microsoft Learn

📘What is Azure AI Content Safety?Service overview, features & limits 🏷️Harm categoriesCategories & severity levels (0–7) 🛑Prompt Shields conceptsJailbreak & indirect attack detection 🎯Groundedness detectionKeep LLM responses grounded ⚡Quickstart: Analyze textREST & SDK getting started 🎛️Content Safety StudioExplore & export code, no-code start

Reference APIs & SDKs cited in the doc: Content Safety REST API, Python SDK, C# SDK, Java & JavaScript SDKs, Azure PowerShell & CLI. [Verify: confirm exact URLs against the latest Learn documentation.]

DCSA Customer Briefing · Thank you

Leave them with the overview and quickstart. All resource titles are named in the source doc; flag that exact URLs should be verified against current Learn before sending. Offer to share this deck + a tailored PoC plan as follow-up.

Azure AIContent Safety

Generative AI created a new attack surface

A cloud AI service that detects harmful content

One service, six ways to keep AI safe

Analyze Text

Analyze Image

Prompt Shields

Groundedness Detection

Protected Material

Custom Categories

The safety layer between your users and your model

Ingest

Send

Analyze

Decide

Act

🧪 Try it: severity threshold playground

📊 Category & severity result

From zero to endpoint in five steps

Have an Azure subscription

Create a Content Safety resource

Choose resource group, region & tier

Grab your key & endpoint

Explore in Content Safety Studio

A few lines to embed the safety layer

Meets developers where they are

Social & UGC moderation at scale

Social / Messaging Platform

GenAI customer-service chatbot filtering

Enterprise Support Assistant

In-game chat & artifact safety

Gaming Platform

Start free, scale on standard

Safe by design, secure by default

Identity & Access

Data Protection

Responsible AI

Coverage & Limits

What to say — and how to answer back

Your path to production

Free tier

Explore Studio

Guided PoC

Scale on S0

Learn more — straight from Microsoft Learn

Speaker Notes

Azure AI
Content Safety