Skip to content
AI Security

Interview question · Adversarial ML & Model Privacy

How do adversarial images attack vision-language models, how is that different from text in images, and does preprocessing help?

Two different attacks hide in 'malicious images'. One writes instructions a model can read; the other optimises pixels to push the model's internal representation. They need different tests and different defences.

hardNewPreparation tags: Google DeepMind, OpenAI, AnthropicUpdated Nov 2026

Two different attacks hide in 'malicious images'. One writes instructions a model can read; the other optimises pixels to push the model's internal representation. They need different tests and different defences.

Make your first attempt

What would you say before reading the answer?

  1. 01 State the assumptions your answer depends on.
  2. 02 Explain the mechanism and choose a control.
  3. 03 Name a condition that would change your decision.

The worked answer includes an approach, scoring criteria and follow-up questions. Compare the reasoning with your own.

10 answers per topic instead of 5, plus saved progress and bookmarks · costs nothingor unlock all 529 remaining answers · $25

Every worked answer in all 17 topics, every lesson in every course, the full concept curriculum, the complete practice-test bank, and the downloadable security field guides.

6 months · One payment · No auto-renewal

What if · the follow-ups in this answer

The answer works through 6 follow-ups like an interviewer would

A vision-language model checks uploaded ID documents and extracts names and dates, with its output feeding an automated account approval step.

  1. What are the two mechanisms for steering the model through images?
  2. How do you handle typographic injection?
  3. Does preprocessing such as JPEG compression stop perturbations?
  4. What if the extracted fields decide account approval automatically?
  5. How would you test this pipeline?
  6. What design would you recommend?

Each rung has a nudge, a model answer, the principle that carries over and the weak answer to avoid. Premium opens every answer and ladder.

UP NEXT ON YOUR JOURNEY
Next in this trackHow would you test whether a model update increased memorisation, and does a leak that needs thousands of queries matter?Next in this trackHow would you red-team a content moderation classifier for evasion, and report results that are useful to the team that owns it?Next in this trackWhat are the main ways attackers 'break' a machine learning model, and which attack fits which level of attacker access?
DISCUSSION · 0

No comments yet. Share a correction or a different way to answer it.