An AI Offensive Security Researcher investigates how AI systems can be attacked and how AI can be used to attack other systems. The first half of the role looks for new classes of weakness in models, agents and the platforms that run them. The second half studies the offensive cyber capabilities of models themselves, such as whether an agent can find and exploit vulnerabilities, and what that means for safety and deployment decisions.
Organisations hire this role when they need original research rather than testing against a known checklist. Model developers need to understand what their systems could do in an attacker's hands before release. Security vendors need to anticipate how attackers will use AI. Enterprises deploying agents need someone who can think several steps ahead of the threats their security teams already know about.
What does an AI Offensive Security Researcher do?
The work is open-ended and driven by hypotheses. An AI Offensive Security Researcher picks a question, such as whether an agent framework can be manipulated into exfiltrating credentials or whether a model can complete a realistic intrusion task, then builds the environment, tooling and experiments to answer it.
- Researching new attack classes against AI agents, including tool misuse, cross-agent injection, memory poisoning and abuse of delegated permissions
- Designing and running dangerous capability evaluations for cyber offence, such as vulnerability discovery, exploit development and lateral movement tasks
- Building realistic test ranges, capture-the-flag style challenges and sandboxed environments for evaluating model behaviour
- Developing AI-assisted offensive tooling for internal red teams and studying how it changes attacker capability
- Analysing the security of AI supply chains, from open-weight model distribution to plug-in ecosystems
- Translating findings into threat models, mitigations and input for release and deployment policy
- Publishing research and presenting at security and ML venues where appropriate
Key skills and backgrounds
Most candidates are experienced offensive security professionals with a research track record. Typical backgrounds include exploit development, vulnerability research, red-team operations and capture-the-flag competition, combined with a solid working knowledge of modern ML systems. Some come from ML research and have built serious security skills. For capability evaluation roles, the ability to design fair, reproducible tasks is as important as hacking ability.
Strong signals include original published research, conference talks, released tools and disclosure history. Ask candidates how they decide which research questions are worth pursuing and how they know an experiment has answered one.
- Exploit development, reverse engineering and network intrusion techniques
- Practical knowledge of LLM agents, tool calling, orchestration frameworks and sandboxing
- Evaluation design for agentic tasks, including scoring, contamination control and reproducibility
- Python and systems languages for tooling, plus experience building test environments
- Threat modelling with MITRE ATLAS alongside conventional attacker frameworks
- Understanding of responsible disclosure and how to handle dual-use findings
- Clear writing for technical peers and for policy and leadership audiences
Where AI Offensive Security Researchers work
Frontier model labs hire offensive security researchers to evaluate cyber capabilities and to secure their own agents and infrastructure. Government AI safety and security institutes employ them to assess models independently. Cyber security vendors hire them to research AI-enabled threats and build defensive products. Specialist AI security start-ups need them to find the attacks their products defend against, and research-focused consultancies and defence organisations also recruit for the role.
AI Offensive Security Researcher vs related roles
This role overlaps with several neighbours, and the boundaries depend on the team.
- An AI Vulnerability Researcher concentrates on finding specific software flaws in AI frameworks, libraries and infrastructure, while an AI Offensive Security Researcher works across broader attack classes and capability questions.
- An AI Bug Bounty Hunter reports individual issues in live products under a programme's rules, whereas an offensive security researcher sets their own research agenda.
- An AI Red Team Engineer applies known and new attacks to a specific organisation's systems on a regular cycle, with less emphasis on publishing or capability evaluation.
Salary and market data
Pay for AI offensive security researchers reflects seniority, research reputation and the type of employer, with labs, vendors and public bodies each benchmarking differently. Our GenAI and agentic systems salary guide gives current ranges, the AI and deep tech salary guides cover related disciplines, and the deep tech salary calculator helps you test a package.
Hiring AI Offensive Security Researchers with DeepRec
DeepRec recruits security researchers for teams working on GenAI and agentic systems across Europe and the US, from offices in London, Berlin, Zurich, Dublin and Boston. Our consultants screen technically, discussing candidates' research, tooling and approach to dual-use findings before introducing them. Many researchers in this field are well known in their communities and not on job boards, so we search through networks and referrals. We hire on a permanent, contract, retained and executive search basis. To start a search, go to Looking to Hire.