Skip to content
Cheat sheets

AI-103 · Domain 3: Implement computer vision solutions · 13% of the exam

Task 3.3: Implement responsible AI for multimodal content

Filtering unsafe images, defending against prompt injection hidden in image text, and enforcing visual policies such as provenance marks, prohibited symbols and brand rules.

Study it

  • Safety for visual content: filters, injected text and visual policies

    Partly covered by: Guardrails

Sample questions

Try each one before opening the answer. Every option is explained, with the Microsoft documentation page that proves it.

Question 1 · choose 1

A browser-testing agent uses the computer use tool. On one page, text in the screenshot tells the model to open the admin settings, and the next computer_call arrives with pending_safety_checks containing malicious_instructions. What should the application do?

  1. ARaise the guardrail's violence blocking level for the testing agent
  2. BHand the action to the end user to confirm before it runs
  3. CRetry the same request until no safety check appears in the call
  4. DAcknowledge the check automatically and run the action
Show the answer and why

  • ARaise the guardrail's violence blocking level for the testing agent

    Incorrect

    The risk is injected instructions in an image, not violent content.

  • BHand the action to the end user to confirm before it runs

    Correct

    When pending safety checks are returned, the guidance is to hand over actions to the end user to confirm the model behaves as intended.

  • CRetry the same request until no safety check appears in the call

    Incorrect

    The adversarial text is still on the page; retrying hides the warning instead of handling it.

  • DAcknowledge the check automatically and run the action

    Incorrect

    Acknowledging without a person reviewing defeats the check that just flagged a possible injection.

The computer use tool checks screenshots for adversarial instructions and flags them as pending safety checks. Treat a flag as a stop: show the proposed action to the user and continue only on their confirmation.

Microsoft documentation

Task 3.3Report a problem with this question

Question 2 · choose 2

A community site has two new visual policy needs. A newly circulating hate symbol must be caught in uploaded images within hours, without training a model. Partners must also be able to verify that the site's marketing images were generated by its Foundry image model. Which two actions meet these needs? (Choose TWO.)

  1. ARaise the blocking level for the hate category to High on uploads
  2. BAdd the symbol's name to a text blocklist on the site's guardrail
  3. CTell the image model in each prompt to draw a small label in a corner
  4. DCreate a custom categories (rapid) incident with sample images
  5. EUse a model that adds Content Credentials; partners run provenance detection
Show the answer and why

  • ARaise the blocking level for the hate category to High on uploads

    Incorrect

    Stricter harm thresholds don't teach the classifier a symbol it has never seen, and they block more legitimate content.

  • BAdd the symbol's name to a text blocklist on the site's guardrail

    Incorrect

    Blocklists match exact text only; they can't recognize the symbol in an image.

  • CTell the image model in each prompt to draw a small label in a corner

    Incorrect

    A visible label requested in a prompt is unreliable and easy to crop; it isn't verifiable provenance.

  • DCreate a custom categories (rapid) incident with sample images

    Correct

    The rapid API matches images through a lightweight object-tracking model and embedding search, with no training step.

  • EUse a model that adds Content Credentials; partners run provenance detection

    Correct

    Supported Foundry image models attach signed C2PA Content Credentials, which the Content Provenance Detection API or website can verify.

Emerging visual incidents call for custom categories (rapid), which work from a few samples. Provenance is a separate job: supported models embed provenance signals that recipients can verify.

Microsoft documentation

Task 3.3Report a problem with this question

Question 3 · choose 1

A fact-checking team runs suspicious images through the Content Provenance Detection API. For one image, no provenance signal is found. How should the team interpret that result?

  1. AIt proves that the image is safe to publish on the web
  2. BIt found no signal; that doesn't prove a person made the image
  3. CIt proves that the image was taken by a camera
  4. DIt proves that the image was edited after it was created
Show the answer and why

  • AIt proves that the image is safe to publish on the web

    Incorrect

    The API doesn't assess safety or harmfulness.

  • BIt found no signal; that doesn't prove a person made the image

    Correct

    A result with no detected provenance doesn't mean the content wasn't AI-generated; the API doesn't identify all AI content.

  • CIt proves that the image was taken by a camera

    Incorrect

    The API isn't a standalone determination of authenticity.

  • DIt proves that the image was edited after it was created

    Incorrect

    No signal says nothing about edits.

Provenance signals are evidence, not verdicts. Combine them with other checks and editorial judgment before deciding about an image.

Microsoft documentation

Task 3.3Report a problem with this question

Question 4 · choose 1

A team is setting up an agent with the computer use tool to test web forms. Where should the agent's desktop environment run?

  1. AOn an administrator's laptop, which has saved passwords in it
  2. BOn the production database server, for realistic test data
  3. COn a VM with no access to sensitive data or critical systems
  4. DOn any machine, since the safety checks block all attacks
Show the answer and why

  • AOn an administrator's laptop, which has saved passwords in it

    Incorrect

    Injected instructions could steer the agent to those credentials.

  • BOn the production database server, for realistic test data

    Incorrect

    That exposes critical resources to model mistakes and attacks.

  • COn a VM with no access to sensitive data or critical systems

    Correct

    The guidance is to use computer use on VMs with no access to sensitive data or critical resources.

  • DOn any machine, since the safety checks block all attacks

    Incorrect

    Safety checks help but don't remove the risk.

Computer use carries real security and privacy risk. Run it in an isolated environment and give it only what the task needs.

Microsoft documentation

Task 3.3Report a problem with this question

Question 5 · choose 1

A chat platform moderates animated GIF stickers with Azure AI Content Safety multimodal analysis. A tester hides a harmful image in a later frame, and it isn't flagged. What explains this?

  1. AGIF files are always rejected by the multimodal service
  2. BMultimodal analysis checks the text in requests only
  3. CThe severity threshold was set to Off for all image content
  4. DFor animated formats, the service analyzes the first frame only
Show the answer and why

  • AGIF files are always rejected by the multimodal service

    Incorrect

    The GIF was analyzed; only its first frame was used.

  • BMultimodal analysis checks the text in requests only

    Incorrect

    The image content was analyzed; the miss comes from which frame was used.

  • CThe severity threshold was set to Off for all image content

    Incorrect

    The miss comes from frame selection, not the threshold.

  • DFor animated formats, the service analyzes the first frame only

    Correct

    If the format is animated, the service extracts the first frame for the analysis.

Know what a moderation API actually inspects. For animated content, extract and check the frames yourself, or convert to formats the policy can evaluate fully.

Microsoft documentation

Task 3.3Report a problem with this question

Question 6 · choose 1

A children's book publisher using a GPT-image model finds that requests for photorealistic images of children are refused. What does the image generation guidance say about this?

  1. AThe refusal means the deployment is out of quota
  2. BThe model can't draw people of any age in any style
  3. CPhotorealistic images of minors are blocked by default
  4. DThe request failed because PNG output was selected
Show the answer and why

  • AThe refusal means the deployment is out of quota

    Incorrect

    Quota problems return 429 errors, not content refusals.

  • BThe model can't draw people of any age in any style

    Incorrect

    The restriction is specific to photorealistic minors.

  • CPhotorealistic images of minors are blocked by default

    Correct

    That capability is blocked by default and can be requested; enterprise customers are approved automatically.

  • DThe request failed because PNG output was selected

    Incorrect

    Output format isn't related to this restriction.

Image generation includes built-in responsible AI protections. Some restrictions can be changed through an access request; design around the ones that can't.

Microsoft documentation

Task 3.3Report a problem with this question

Question 7 · choose 1

A brand team plans to promise customers that every AI image it publishes carries Microsoft provenance signals. It is evaluating several image models in Foundry. What should it check first?

  1. AWhether each model is the cheapest per generated image
  2. BWhether each model is on the provenance support list
  3. CWhether each model can return its images as PNG files
  4. DWhether each model supports the n parameter in requests
Show the answer and why

  • AWhether each model is the cheapest per generated image

    Incorrect

    Price doesn't determine provenance coverage.

  • BWhether each model is on the provenance support list

    Correct

    Provenance capabilities cover selected models; support varies by modality, model family and deployment type.

  • CWhether each model can return its images as PNG files

    Incorrect

    Output format doesn't decide provenance support.

  • DWhether each model supports the n parameter in requests

    Incorrect

    The number of images per call is unrelated.

Provenance is available for selected Foundry models only; verify coverage for each model before making promises about published content.

Microsoft documentation

Task 3.3Report a problem with this question

Practise domain 3 →Practise all domains →