PhD Research Fellowship in Safety and Security Evaluation of Deployed AI Systems

Project / Job description A 3-year PhD Research Fellowship is available at Simula Metropolitan Center for Digital Engineering (SimulaMet), within our research effort on the security and safety of AI systems. Most security and safety work on AI happens where models are built. We work where they are used. As AI systems move into hospitals, public administration, and other consequential settings, the institutions that deploy them, and the regulators and auditors who oversee them, need to establish whether a given system is safe to use, independently of the vendor that built it. Today, they largely cannot. Closing that gap is the purpose of this position, and it grows more urgent as accountability shifts toward the organisations that deploy and audit AI, not only those that train it. The PhD researcher will develop methods and tools for evaluating, red-teaming, and monitoring AI systems as they are actually deployed, large language models today, and increasingly agentic and multimodal systems. The work pursues two goals at once: A science of failure. Not counting failures, but explaining why and when systems fail, so that findings generalise beyond the single system tested. This means treating a model as one part of a larger socio-technical system, evaluating data, training, deployment, and feedback continuously rather than once, and weighting the catastrophic 1% over the comfortable 99% average. Open, vendor-independent tools. Practical, local-first instruments that any regulator, auditor, hospital, or public agency can run without commercial dependencies and without access to the vendor of the system under test. You would build on and extend an existing, openly released toolset: SimpleAudit (scenario-based safety auditing; a verified Digital Public Good), swarm-attack (adversarial swarm red-teaming), and the Moltbook Observatory (passive monitoring of deployed agents). The precise direction will be shaped together with you and the supervisory team, and may emphasise scenario-based auditing, adversarial and swarm red-teaming, agentic-system security, continuous monitoring, multilingual and national-language deployment, or the methodology of evaluation itself. A defining strength of the group is direct collaboration with the institutions where failure is costly, including partners across Norwegian public health and the public sector, which grounds the work in real deployment contexts rather than synthetic benchmarks. The purpose of a PhD fellowship is to qualify the candidate for work in positions that require a high level of scientific competence, leading to the award of a doctoral degree (PhD). The successful candidate will be admitted to a doctoral (PhD) programme at OsloMet - Oslo Metropolitan University. A career development plan will be drawn up for the fellow, and SimulaMet is responsible for following up on this plan and ensuring access to supervision and career guidance throughout the period. Research stays abroad and presentation of results at international and national conferences are facilitated and strongly encouraged, and training and support for academic writing are provided throughout the fellowship. Candidate Profile We will consider candidates who have (or will have completed before the start of the position) a Master's degree (or equivalent) in Computer Science, Applied Mathematics, or another relevant discipline. The applicant is required to demonstrate how the degree corresponds to the profile of the post. Other qualifications include: A strong background in machine learning, data science, benchmarking and evaluation of AI systems, AI security and safety, AI alignment or a closely related area. Research experience, demonstrated for example through a strong Master's thesis, and ideally through publicly released code or publications. Proficiency in a high-level programming language (e.g., Python). Excellent oral and written communication skills in English. An open and collaborative mindset, with the ability to…