AI Safety & Governance Engineer (SAGE)

What matters most

• Pre-delivery safety gates

  The role runs automated LLM safety tests in GitLab CI/CD as a mandatory gate before delivery to project sponsors. For preparation, outline how you would build maintainable Python tests with pytest or unittest and make failures actionable.

• Adversarial testing

  The posting calls for red teaming prompt injection, jailbreaking, data leakage, RAG poisoning and multi-turn attacks. Prepare to explain how you would construct reproducible tests and distinguish a demonstrated vulnerability from an unsuccessful attack attempt.

• Secure architecture review

  The engineer reviews agentic, multimodal and RAG systems for vulnerabilities and recommends secure designs. Prepare to reason through trust boundaries, API integrations and where untrusted content could influence system behaviour.

• Actionable findings and governance evidence

  Findings must include severity, reproduction steps and remediation guidance; the role also supports compliance with the named AI safety frameworks. Prepare to connect a technical finding to a practical fix and clear supporting evidence, rather than reciting framework names.

• Apprentice mentoring and knowledge sharing

  The role trains AIAP apprentices during deep-skilling and project work and contributes talks, workshops and documentation. Prepare an approach for explaining responsible AI engineering practices and reviewing others’ work constructively.

Likely challenges to think through

• Reliable gates for variable LLM behaviour

  Possible challenge, not a stated employer problem: automated safety checks may need to catch meaningful failures without treating every variation in model output as a release blocker. Prepare to discuss test design, reproducibility and how to investigate a failing CI result.

• Coverage across different system designs

  Possible challenge: work across industry projects and agentic, multimodal and RAG architectures could make one fixed attack suite insufficient. Prepare to show how you would adapt threat scenarios to each system’s data flows and trust boundaries while retaining repeatable baseline tests.

• Turning security findings into usable decisions

  Possible challenge: a technically valid exploit may be hard to prioritise or remedy without context, especially when documenting compliance evidence. Prepare to explain how severity, reproduction steps, impact and remediation guidance would help a project team act on a finding.

Topics to prepare

• LLM, RAG and API trust boundaries

  The posting requires working knowledge of LLM architectures, prompt engineering, RAG implementations and API integration. As revision, sketch how a user request, retrieved content and API response reach an LLM; mark where lower-trust content could influence instructions.

• LLM-specific attack mechanisms

  Red teaming in the role covers prompt injection, jailbreaking, data leakage, RAG poisoning and multi-turn attacks. For each, write a short example of the attacker's entry point, intended outcome and observable failure.

• Security design for agentic and multimodal systems

  The role reviews agentic, multimodal and RAG architectures and recommends secure designs. As revision, compare where an agent might gain authority to take an action, where multimodal input enters, and what boundaries a design review should examine.

• AI safety and governance frameworks

  The posting names ISO 42001, NIST AI RMF, OWASP LLM Top 10, MITRE ATLAS and the Singapore Model AI Governance Framework. Make a one-page comparison of how each could inform risk identification, testing or compliance evidence; do not treat them as interchangeable.

• Pre-delivery safety gates

  SAGE runs automated LLM safety tests in GitLab CI/CD as a mandatory gate before delivery. As revision, diagram how a test result, failure and supporting evidence would pass through a pipeline to a delivery decision. Specific pass criteria are not stated in the posting.

Skills to revise

• Design adversarial test cases

  The role conducts red teaming and adversarial testing. Practise writing reproducible test cases for prompt injection, RAG poisoning and a multi-turn data-leakage attempt, with expected safe behaviour and a way to record the observed result.

• Implement Python safety tests

  Proficient, maintainable Python and experience with pytest or unittest are required. As practice, write a small parameterised suite for example LLM responses, including clear assertions and failure output; this is a preparation exercise, not a specified employer test suite.

• Connect tests to CI/CD

  The posting specifies GitLab CI/CD safety-testing pipelines and familiarity with Git, Docker and GitLab or GitHub pipelines. Practise outlining a GitLab pipeline that runs the Python tests and blocks progression on failure, then explain how you would inspect its results.

• Review an architecture for vulnerabilities

  Architecture review and secure-design recommendations are explicit duties. Take a hypothetical agentic RAG diagram, trace retrieved content and available actions, then write two vulnerability findings with proposed design changes. The diagram is a practice scenario, not a stated project.

• Write findings and compliance evidence

  The role calls for severity ratings, reproduction steps, remediation guidance and strong technical writing. Turn one practice test failure into a concise vulnerability report, then identify which test records could support a relevant governance review.

• Explain responsible AI testing to an apprentice

  The posting includes mentoring AIAP apprentices and contributing talks, workshops and documentation. Practise a brief walkthrough of one prompt-injection test, explaining its trust boundary, result and remediation in terms an apprentice could use.

• Compile a role-evidence sheet

  The posting asks for a degree, 2–4 years of work experience, and evidence of LLM, security-testing and software-engineering skills. Prepare a one-page index of relevant, verifiable work; distinguish direct experience from areas you have only studied.

• Rehearse an LLM project walkthrough

  The role requires knowledge of LLM architectures, prompting, RAG and API integration. Choose a shareable project, or a clearly labelled illustrative example, and explain its data flow, design decisions, risks and your contribution.

• Sketch a pre-delivery testing gate

  SAGE runs automated LLM safety tests in GitLab CI/CD before delivery. Prepare a small Python test example using pytest or unittest and a pipeline sketch showing when tests run, what constitutes a failure and what evidence is retained.

• Prepare an adversarial test report

  The posting names prompt injection, jailbreaking, data leakage, RAG poisoning and multi-turn attacks. Draft a safe example test plan and one sample finding with a severity rationale, reproduction steps and remediation guidance.

• Practise a secure architecture review

  The role reviews agentic, multimodal and RAG systems. Annotate an illustrative architecture with trust boundaries and plausible vulnerabilities, then explain design recommendations and their trade-offs.

• Map findings to governance evidence

  SAGE supports ISO 42001, NIST AI RMF, OWASP LLM Top 10, MITRE ATLAS and Singapore’s Model AI Governance Framework. Prepare a brief example showing how a test result or finding could be documented against relevant guidance; do not present it as proof of certification or compliance.

• Plan a short apprentice teaching exercise

  The engineer mentors AIAP apprentices and contributes talks, workshops and documentation. Outline a 10-minute explanation of an LLM vulnerability, a hands-on exercise and the feedback you would give.

• Write questions about SAGE’s working practices

  Ask how the team selects tests across 100 Experiments projects, handles pre-delivery gate failures, reviews remediation and divides mentoring time with project work. These are questions to clarify, not assumptions about its processes.