Careers / Prep packs

Mid level

AI Safety and Evaluation interview questions and prep

Mid-level AI safety evaluators test large language models and AI agents for vulnerabilities, misuse and harmful outputs through red-teaming, prompt injection, jailbreak attempts, and measurement of safety guardrails. Interviews assess your ability to design realistic attack scenarios, reason about model behaviour under adversarial conditions, spot unsafe system architecture, and balance thorough evaluation with pract

  • 30 questions
  • 8 skills
  • 10 labs

Skills this role needs

  • Prompt injection attack designCore skill for testing whether models correctly distinguish user input from system instructions and confidential data.
  • Jailbreak and refusal suppression techniquesEssential to probe the limits of safety filters, roleplay constraints, and instruction hierarchy in deployed models.
  • Agent and tool-use hijackingAgents with function calling and external tools introduce new attack surfaces; you must evaluate whether tool selection, argument handling, and output processing are robust.
  • Data exfiltration and leakage vectorsModels can leak training data, system prompts, context, and user information through direct extraction, encoding channels, or indirect inference; identifying these paths is core evaluation work.
  • Unsafe output sinks and downstream integrationModel safety means nothing if output is rendered as code, SQL, HTML, or OS commands without sanitisation; you must map how outputs flow into downstream systems.
  • Red-team scenario design and measurementEvaluation requires systematic, reproducible attack templates and metrics to compare model robustness across versions and deployments.
  • AI governance and policy interpretationSafety testing must align with organisational risk tolerance, regulatory requirements, and responsible AI frameworks; you translate policy into concrete test cases.
  • Common web and API vulnerabilitiesAI safety integrates with application security; injection, CSRF, SSRF, authentication bypass, and cache poisoning remain relevant even when the attack surface is a language model.

Labs to practise

Hands-on labs on this site that match the work in this role. They run in the browser.

Courses that help

Tutorials to read

30 interview questions

Answer in your own words first, then open what a strong answer covers. Sign in to practise with AI feedback.

  1. You are evaluating a customer support chatbot. Walk through how you would test for prompt injection via user messages. What payloads would you try, and how would you know if the bot's safety guardrails had failed?

    What a strong answer covers
    • prompt injection attack patterns
    • distinguishing user input from system instructions
    • defining success criteria for a failed safety filter
  2. A model is instructed never to reveal its system prompt. Design a red-team plan to extract it. What techniques would you try in what order, and why?

    What a strong answer covers
    • system prompt extraction methods
    • iterative attack escalation
    • understanding model attention and context windows
  3. An AI agent can call a web search tool, a calculator, and a code interpreter. Describe two ways an attacker could hijack the agent's behaviour by crafting inputs or poisoning tool outputs. What guardrail would prevent each?

    What a strong answer covers
    • agent tool-call injection
    • tool output hijacking
    • agent permission and output validation design
  4. You find that a model's output is rendered directly as HTML in a web page without sanitisation. Craft a prompt that would cause XSS. What does this tell you about the product's safety architecture?

    What a strong answer covers
    • unsafe output sinks
    • downstream integration risks
    • separation of model output from rendering logic
  5. A company fine-tunes a model on proprietary customer data. Design a membership inference attack to test whether the model has memorised specific records. What would constitute a vulnerability, and how would you measure severity?

    What a strong answer covers
    • training data privacy risks
    • membership inference methodology
    • privacy metrics and reporting
  6. The model refuses a harmful request with 'I can't help with that.' Outline a jailbreak strategy using roleplay, encoding, or multi-turn escalation. At each step, explain why the model might fail to detect the bypass.

    What a strong answer covers
    • refusal suppression techniques
    • reasoning about model instruction hierarchy
    • multi-turn context and instruction confusion
  7. You are designing a safety test suite for a model that will be deployed in a high-stakes domain (medical, financial, or legal). What categories of risks must your red-team cover, and how would you prioritise them?

    What a strong answer covers
    • risk-based evaluation planning
    • domain-specific harm models
    • stakeholder communication on safety findings
  8. A model is asked to summarise a user-uploaded PDF. An attacker embeds a prompt injection in the PDF metadata or hidden text. How would you test for this vulnerability? What output would indicate a successful attack?

    What a strong answer covers
    • indirect prompt injection via documents
    • untrusted input validation
    • multimodal attack surfaces
  9. Explain the difference between direct prompt injection, indirect prompt injection, and instruction conflict. Give one realistic example of each in a deployed AI product.

    What a strong answer covers
    • prompt injection taxonomy
    • real-world deployment contexts
    • vulnerability surface mapping
  10. A model is used to generate SQL queries from natural language. Design a text-to-SQL injection attack. What safeguard would you recommend, and why might a simple output filter not be enough?

    What a strong answer covers
    • query generation vulnerabilities
    • semantic vs. syntactic defense
    • query parameterisation and validation
  11. You are red-teaming a retrieval-augmented generation (RAG) system. Describe three attack vectors that target the ingestion, retrieval, or ranking stages. How would each bypass the model's safety training?

    What a strong answer covers
    • RAG poisoning and embedding attacks
    • knowledge base integrity
    • ranking manipulation and relevance spoofing
  12. An AI coding assistant generates code that is committed to a repository. What security flaws might the generated code contain, and how would you audit them? Name at least three categories.

    What a strong answer covers
    • AI-generated code vulnerabilities
    • false confidence and test gaps
    • dependency and supply-chain risks
  13. A multi-agent system uses persistent memory to store conversation history. Explain how memory poisoning could hijack the agent's future behaviour. What governance control would you put in place?

    What a strong answer covers
    • persistent state vulnerabilities
    • long-horizon instruction injection
    • audit and oversight mechanisms
  14. You are evaluating a model's compliance with a responsible AI policy that bans illegal content, hate speech, and violence. How would you operationalise these abstract terms into concrete red-team test cases?

    What a strong answer covers
    • policy operationalisation
    • harm definition and taxonomy
    • borderline case handling and measurement
  15. A model is used to classify phishing emails. An attacker crafts an adversarial email that looks innocuous to humans but fools the model. How would you systematically search for such evasion examples? What does this vulnerability reveal about the model?

    What a strong answer covers
    • adversarial evasion testing
    • robustness vs. accuracy tradeoffs
    • distribution shift and generalisation
  16. The model is integrated with a human review loop: users request an action, the model drafts a summary of its reasoning, and a human approves or rejects it. Describe how an attacker could spoof the approval summary to bypass the human reviewer.

    What a strong answer covers
    • human-in-the-loop bypass
    • trust boundary violations
    • transparency and interpretability gaps
  17. An LLM-powered agent can read files from a shared directory. Outline a path traversal or sandbox escape attack. What least-privilege controls would mitigate this risk?

    What a strong answer covers
    • code execution and sandbox vulnerabilities
    • file system access control
    • capability restriction and blast radius
  18. You discover that two separate users of the same AI platform share an embedding cache. Explain how this could leak information from one user's documents to another. How would you test for this vulnerability?

    What a strong answer covers
    • multi-tenant isolation failures
    • cache-based side channels
    • information leakage across trust boundaries
  19. A model is asked to identify the author of an anonymous text by comparing it to a set of known writing samples. What privacy risks does this create, and how would you measure whether the model is vulnerable to membership inference or training data extraction?

    What a strong answer covers
    • privacy risks in comparison and matching tasks
    • privacy measurement techniques
    • regulatory and ethical considerations
  20. An attacker sends a request that is relayed through multiple AI agents in sequence. Each agent transforms the input slightly. Explain how this could bypass a single-agent safety filter. How would you test for this 'whisper down the lane' vulnerability?

    What a strong answer covers
    • multi-agent attack chains
    • instruction drift and transformation
    • system-level safety evaluation
  21. A model uses tools to fetch URLs and summarise their contents. Design an SSRF attack that tricks the model into accessing an internal service. What guardrails would prevent this?

    What a strong answer covers
    • SSRF via AI-driven HTTP requests
    • tool parameter validation
    • network segmentation and access control
  22. You are building a red-team report to present to non-technical stakeholders. How would you communicate the severity of a prompt injection vulnerability that leaks the system prompt but not user data? What risk metrics would you include?

    What a strong answer covers
    • risk communication and severity assessment
    • stakeholder translation
    • business impact reasoning
  23. A model is fine-tuned on a small dataset that includes sensitive information. Propose a test to detect whether the model has overfit and memorised specific records. How would you balance privacy with model quality assessment?

    What a strong answer covers
    • overfitting and memorisation detection
    • fine-tuning security risks
    • privacy-utility tradeoffs
  24. An attacker uses a low-resource language (one the model rarely saw during training) to submit a jailbreak prompt. Explain why this might be more effective than English. How would you test for multilingual vulnerabilities?

    What a strong answer covers
    • multilingual safety gaps
    • training data bias and coverage
    • cross-lingual attack transfer
  25. A model is instructed to output JSON in a specific schema. An attacker crafts a prompt that makes the model output valid JSON that violates the real business logic (e.g., a negative quantity or a future date that breaks a workflow). How would you test for this? Why is schema validation alone insufficient?

    What a strong answer covers
    • output schema bypass
    • semantic constraint violation
    • business logic integrity
  26. You are designing a red-team exercise for a new AI product launch. Outline a 2-week evaluation plan. What would you test in week 1 vs. week 2? How would you decide when to stop testing?

    What a strong answer covers
    • evaluation planning and scoping
    • test prioritisation
    • risk-based stopping criteria
  27. A model generates code for a SQL query, an OS command, or an HTML email. For each output type, name the top two security risks and the validation technique you would use to mitigate it.

    What a strong answer covers
    • output risk taxonomy by sink type
    • code generation vulnerabilities
    • context-specific validation strategies
  28. An AI agent is used to review and merge pull requests. Describe two ways an attacker could use a poisoned GitHub issue, comment, or PR to hijack the agent's code review logic. How would you test for this vulnerability?

    What a strong answer covers
    • indirect prompt injection via platform data
    • DevOps and CI/CD agent risks
    • code integrity and approval bypass
  29. You find that a model's refusal message is itself vulnerable to prompt injection (e.g., 'I can't help with that [INJECTION HERE]'). Explain why this is dangerous and how you would test for it.

    What a strong answer covers
    • layered prompt injection via error handling
    • failure mode analysis
    • cascading vulnerability chains
  30. A model is used to detect fraudulent transactions. An attacker crafts transactions that are misclassified as legitimate. Explain how this differs from a prompt injection vulnerability. What evaluation methodology would you use to measure robustness?

    What a strong answer covers
    • adversarial examples vs. prompt injection
    • classifier evasion
    • robustness benchmarking and metrics

A 14-day preparation plan

  1. Days 1 to 3Complete OWASP Top 10 for LLM Applications course (5628) and work through foundational prompt injection labs (prompt-injection-1, ai-instruction-conflict-1, ai-refusal-suppression-1). Read the role confusion and agent hijacking labs (5169, 5157) to map how instruction boundaries fail. Document attack patterns in a pers
  2. Days 4 to 7Take Prompt Injection From Zero course (5533) and practice extraction attacks (ai-system-leak-1). Work through data exfiltration and output sink labs (ai-data-exfil-1, ai-llm-output-xss-1). Read MCP Attack Surface (4667) and AI Governance assessment (4452). Build a template for red-team scenario design covering prompt
  3. Days 8 to 11Take Securing AI Agents course (5537) and complete agent-focused labs (ai-agent-hijack-1, ai-tool-output-injection-1). Read tool output injection and confused deputy labs (5158, 5160). Practise designing attacks where tool calls are hijacked or tool results poison the agent's reasoning. Sketch threat models for multi-a
  4. Days 12 to 14Synthesise findings: write a sample red-team report covering three vulnerability categories (prompt injection, agent hijacking, unsafe output sink). Review governance reading (4452) and practise translating vulnerabilities into risk metrics for non-technical audiences. Revise scenario design and prepare to discuss your

These questions are SecurityElites practice material for this kind of role. They are not taken from any employer, and a real interview may differ.