An AI Security Researcher studies how machine learning systems can be attacked, misused or made to fail, and develops methods to prevent it. The role is broader than any single attack type. It spans the model, its training data, the agents and tools built around it, and the processes that decide what gets deployed. The aim is to give engineering and leadership teams a clear, evidence-based picture of risk and a set of defences that hold up under scrutiny.
Organisations hire AI security researchers when they need someone to set direction rather than run tests. A lab preparing more capable models, a company putting agents into customer workflows or a regulated business adopting generative AI all face open questions about threats nobody has fully characterised yet. A researcher defines those threats, tests them properly and turns the results into design choices.
What does an AI Security Researcher do?
The role mixes research, prototyping and advisory work. An AI Security Researcher identifies the threats most relevant to the organisation's systems, runs experiments to understand them, and works with engineers to build mitigations into products and training processes.
- Developing threat models for models, agents, retrieval systems and training pipelines
- Researching attacks including prompt injection, data poisoning, backdoors, model extraction and membership inference
- Designing secure architectures for agentic systems, such as privilege separation, sandboxing and approval controls on tool use
- Evaluating defences under adaptive attack and publishing honest assessments of their limits
- Investigating provenance and integrity methods for models and data, including signing, watermarking and content authenticity
- Building evaluation suites that let teams measure security properties across model versions
- Advising on governance and assurance work linked to the NIST AI RMF and the EU AI Act
Key skills and backgrounds
Most AI security researchers hold a PhD in machine learning, computer security or a closely related field, or have several years of equivalent industry research. Strong candidates have a foot in both worlds. They can read and reproduce ML papers, and they think like an attacker about how a system will be used in practice. Breadth is useful in this role because the threats change as model capabilities and deployment patterns change.
Look for a coherent research agenda, publications or released tools, and a habit of evaluating defences fairly. Ask candidates which of their own results they trust least, and why.
- Strong foundations in deep learning, including training, fine-tuning and evaluation of large models
- Security research methods, threat modelling and attacker reasoning
- Experiment design, reproducibility and statistical care
- Proficiency in Python with PyTorch or JAX, plus enough systems knowledge to test real deployments
- Knowledge of MITRE ATLAS and the OWASP guidance on LLM application risks
- Awareness of privacy attacks and defences, including differential privacy
- Clear writing for research audiences, engineers and decision makers
Where AI Security Researchers work
Foundation model labs employ AI security researchers in dedicated security and safety teams. Large technology companies hire them into research groups that support product security. AI security start-ups need them to build the core methods behind their products. Government AI safety and security institutes, defence research bodies and national cyber agencies recruit them to study risks independently. Universities and non-profit research organisations also run AI security programmes, often in collaboration with industry.
AI Security Researcher vs related roles
AI security research sits close to several research specialisms. These distinctions help when writing a job description.
- An ML Security Research Scientist usually works on more formal, publication-led research into specific attack and defence problems, where an AI Security Researcher often covers a wider range of applied questions.
- An Adversarial Robustness Researcher focuses on how models behave under adversarial or shifted inputs, which is one part of the AI security research agenda.
- An AI Safety Researcher studies harmful behaviour that arises from the model itself, such as misaligned goals or dangerous capabilities, rather than threats from an external attacker.
Salary and market data
Pay for AI security researchers depends on research seniority, publication record and employer type, with labs, start-ups and public bodies all structuring packages differently. See our GenAI and agentic systems salary guide for current ranges, the wider AI and deep tech salary guides for research roles, and the deep tech salary calculator to benchmark an offer.
Hiring AI Security Researchers with DeepRec
DeepRec recruits for AI research and AI security teams across Europe and the US, from offices in London, Berlin, Zurich, Dublin and Boston. Our consultants screen researchers technically, discussing their threat models, experiments and how they evaluated defences, so the shortlist reflects real research depth. Many strong candidates are in academia or established labs and not actively looking, so we search through research networks as well as applications. We work on a permanent, contract, retained and executive search basis. To start a search, visit Looking to Hire.