What an AI security specialist does
AI security is the discipline of protecting systems that use machine learning and large language models (LLMs). In practice the work falls into three areas, and most roles lean heavily on one of them:
- Breaking AI. Adversarial testing, usually called AI red teaming: making an LLM application leak data, follow injected instructions, ignore its guardrails or misuse the tools it has been given.
- Securing AI. Designing AI features safely: threat modelling retrieval-augmented generation (RAG) pipelines and AI agents, limiting tool permissions, building guardrails and evaluation harnesses, and securing the pipelines that train and serve models.
- Governing AI. Deciding which AI is acceptable and proving it is controlled: AI policies, risk and impact assessments, and management systems aligned to standards such as ISO/IEC 42001.
Knowing which of the three you need is the most important decision in the brief. A brilliant AI red teamer is rarely the right person to write your AI governance framework, and the reverse is almost always true.
Signs you need one
- You have an LLM feature that customers or staff can use, especially one that reads internal documents or emails.
- You're giving AI agents tools: the ability to send emails, create tickets, query databases or move money.
- Your board or regulator is asking how AI risk is governed, or you're preparing for ISO/IEC 42001.
- Your security team's pen tests cover the web app around the AI, but nobody tests the AI itself.
The main AI security roles
| Role | Focus | Typical background |
|---|---|---|
| AI red teamer | Adversarial testing of LLM apps and agents | Penetration testing, offensive security |
| LLM security engineer | Building guardrails, evaluations and secure AI features | Application security, software engineering |
| AI security architect | Threat modelling and secure design of AI systems | Security architecture, cloud security |
| MLSecOps engineer | Securing model training, supply chain and serving | ML engineering, DevSecOps |
| AI governance lead | Policy, risk, ISO/IEC 42001 and assurance | GRC, privacy, risk management |
| Head of AI security | Strategy, team and executive reporting | Security leadership with hands-on AI exposure |
Skills and frameworks to test for
Four reference points come up again and again. A credible candidate should know the ones relevant to the role well enough to apply them, not just name them.
- OWASP Top 10 for LLM Applications: the common vulnerability classes, including prompt injection, sensitive information disclosure and excessive agency.
- MITRE ATLAS: a knowledge base of real adversary tactics and techniques against AI systems, the AI counterpart of MITRE ATT&CK.
- ISO/IEC 42001: the international standard for an AI management system, central to governance roles.
- NIST AI Risk Management Framework: a widely used structure for identifying and managing AI risk.
In Australia, add awareness of privacy obligations for automated decision-making, which require organisations covered by the Privacy Act to explain in their privacy policy, from 10 December 2026, when computer programs substantially help make decisions that significantly affect people. For government work, add the Australian Government's policy and guidance on responsible use of AI.
Where AI security specialists come from
The discipline is young, so almost nobody has ten years of "AI security" on their CV. The strongest candidates usually come from:
- Offensive security: pen testers and red teamers who have started testing chatbots and agents.
- Application security: engineers who threat model and review AI features as part of the product lifecycle.
- ML engineering: people who built the models and pipelines and became the de facto security owners.
- GRC and privacy: risk professionals who led their organisation's first AI policy or impact assessments.
That is why title-based searches fail. Search for evidence of the skills, such as an AI red-team engagement, a guardrail they built, or an AI policy they wrote, not for the words "AI security".
Interview questions that work
Scenario questions beat definition questions. Anyone can recite the OWASP list; fewer people can apply it under pressure.
- "An internal assistant can read SharePoint and create support tickets. How would you red-team it before launch?" Look for indirect prompt injection via documents, data exfiltration through tool calls, permission boundaries and a plan to retest.
- "Walk me through a vulnerability you found in an AI system, how you proved it and what fix you recommended." Look for specifics: the system, the technique, how reproducible it was, and a fix that changed the design rather than a filter on one prompt.
- "A product team wants to let an agent issue refunds. What are your launch conditions?" Look for human approval thresholds, least-privilege tools, logging and evaluation before go-live.
- "How would you measure whether a guardrail actually works?" Look for evaluation sets, attack success rates before and after, and regression testing as models change.
- For governance roles: "How would you build an AI risk register for a company with forty AI use cases?" Look for an inventory, risk tiers, impact assessments, ownership and review cycles.
Tip: ask one follow-up on every answer: "what exactly did you do, and how did you know it worked?" It is the fastest way to separate lived experience from rehearsed answers. Critical Hive does this automatically in its AI interviews.
Red flags
- Talks about "jailbreaks" but can't explain indirect prompt injection or tool misuse.
- Treats a keyword blocklist as a security control.
- No way of measuring whether a fix worked.
- Answers that are polished, generic and identical in structure, especially in remote interviews. Check for AI-assisted answers and identity issues; see our guide to interview fraud and deepfake candidates.
Your hiring options
- Permanent hire when AI is core to your product or you have a sustained pipeline of AI features.
- Contract for a launch, an assessment or a time-boxed red-team program.
- Advisory first when you need answers before you can write the job description; our AI security advisory can test and govern your AI while you hire.
Critical AI recruits AI security specialists across Australia and assesses every shortlisted candidate with a structured AI interview and a practitioner review. Brief us on a role.