PhD Research Fellow Position in Knowledge-Grounded LLM Agents for Defensive Cyber Operations (ref 309937)

PhD Research Fellow Position in Knowledge-Grounded LLM Agents for Defensive Cyber Operations (ref 309937)

Evidence as of 17 October 2026. University of Oslo (UiO), Department of Informatics, Digital Security Research Group; Norway. Recruitment reference 309937.

Quick Decision

This project asks how a language-model agent can turn fragmented security observations into traceable evidence and controlled defensive decisions. It combines semantic knowledge graphs, symbolic reasoning and LLM tool use, with FFI co-supervision. Recent joint papers provide concrete starting points in alert investigation and contextualisation. The strongest matching preparation is substantial programming plus computer-science/AI foundations, with cybersecurity knowledge or a demonstrated motivation to apply AI in the field.

Promising with material uncertainties. The official recruitment advertisement in the retained evidence confirms a paid three-year doctorate, starting as soon as possible and no later than 1 May 2027, with 30 November 2026 deadline. A fourth year is only a possibility. Crucially, the job explicitly requires security eligibility and may involve background checks and national-security stakeholders. It states no blanket Chinese-nationality exclusion, but feasibility for a China-resident applicant remains unconfirmed. Establish the project's actual assessment/access requirements before investing heavily in an application or presenting it as unrestricted international recruitment.

This report reuses the saved material; the current application status has not been checked.

Research Problem and Doctoral Work

Evidence-grounded investigation is the central question

The advertised work transforms heterogeneous cyber intelligence, telemetry and observations into structured knowledge, then combines semantic graphs, symbolic reasoning, language models and autonomous agents for analysis, situational awareness, decisions and response. Trustworthiness, reliability, interpretability and security are central; adversarial manipulation, uncertainty, provenance and validation of generated actions may also be explored.

The central research challenge is determining which observations support a conclusion, which information is missing, whether a proposed action is permitted and how errors propagate across tools. A plausible doctoral contribution could compare representations/reasoning constraints under common evidence and measure reliable investigation rather than fluent explanations. This is an interpretation of the research problem, not a promised work package or guaranteed autonomous deployment.

The team’s alert-investigation paper is a concrete precursor: constrained access to log evidence supports an LLM’s investigation. The graph-contextualisation preprint groups related alerts before comparison with previous incidents. Together they suggest a progression from observations to context to decisions, while leaving open data quality, generalisation, provenance and operational control.

Possible experiments and realistic limits

Useful research angles to discuss include:

• Compare a plain LLM/retrieval baseline with graph-based grounding and symbolic constraints using the same evidence, temporal splits and explicit missing-data cases.
• Evaluate conclusions with traceable evidence and appropriate abstention, alongside accuracy, investigation cost and analyst effort.
• Assess reliability when retrieved evidence is untrusted, including whether authority limits, decision validation and human oversight contain errors.

These are analyst suggestions, not extra admission conditions. The CACAO 2.0 committee specification offers an example of structured defensive playbooks. Its schema does not itself establish that an action is authorized or correct, and no requirement to use CACAO is advertised.

Independent AgentDojo research at NeurIPS 2024 shows why utility and security must be evaluated together for tools operating on untrusted data. A 2 October 2026 preprint reports poor transfer of injection-detector rankings across benchmarks and argues for testing on actual agent-style tool outputs at low false-positive rates. That recent preprint is emerging external evidence, not settled consensus or UiO output. Neither source is a direct SOC deployment benchmark.

The advertisement does not allocate a dataset, GPU budget, model/API access, industrial deployment or classified environment. Define these early, together with publication rights and an open-data alternative where sensitive access is unavailable.

School, Programme and Field Context

UiO is both the advertised employer and the university doctoral route; Informatics is in its Faculty of Mathematics and Natural Sciences. The official Digital Security unit confirms the organizational home at Gaustadalléen 23 B in Oslo. FFI contributes co-supervision and application context. The sources establish neither an FFI employment contract nor a joint/double degree.

The faculty doctoral regulations provide the training framework: employed faculty fellows are admitted from commencement, then formalize their PhD agreement. The educational component is 30–40 credits, with a progress review after 12–15 months, and thesis/defence in English. Employment selection therefore links to a documented research-training process; exact main/co-supervisor roles and approved plan remain to be agreed. A three-year contract is not a graduation guarantee. Any fourth year involves a different workload arrangement and is not secured by the advertisement.

The ranking-provider page read on 17 October 2026 labels UiO 86th in ARWU 2026 overall and 51–75 in GRAS 2026 Computer Science & Engineering. These are university/subject indicators, not a cybersecurity-lab ranking or evidence of mentoring quality. They complement the more directly relevant joint research and doctoral framework. The page's year and scope are retained rather than mixed with older indexed rankings.

Supervisor and Lab Assessment

Confirmed people and relationships

Vasileios Mavroeidis’s official profile identifies an Associate Professor in Digital Security, working on knowledge representation, evidence/provenance, constrained agents and interoperable cyber defence. It documents MSc/PhD supervision and involvement with ENISA/OASIS. The recruitment contact list uses the broad title Professor; this report follows the more specific staff-profile rank.

Gudmund Grov’s academic homepage identifies him as an FFI Principal Scientist and UiO adjunct Associate Professor, with formal-reasoning experience and doctoral supervision. He and Mavroeidis are both named recruitment contacts. Their exact primary/co-supervisor allocation for this new appointment is not stated. The evidence supports an existing UiO/FFI research connection, without proving guaranteed classified-data access, mentoring frequency or a particular student completion rate.

Recent Research Contributions

Eilertsen, Mavroeidis and Grov: Towards Agentic Investigation of Security Alerts

This is a 2025 IEEE Big Data paper, with an author preprint posted in April 2026. The paper record/abstract describes constrained investigation of log evidence, improving verdicts relative to the same LLM without that workflow in the reported setting. This is relevant to evidence gathering; it does not establish robust production use.

Eckhoff et al.: Neurosymbolic defensive analysis, 2025

The publisher PDF examines several combinations of learned and symbolic methods, including logic-aware detection, alert contextualisation and knowledge-based threat hunting. It extends NeSy 2024 work. The authors explicitly limit their experiments to monitoring/analysis proof of concepts and discuss scalability; this is a foundation for further research, not a validated autonomous SOC.

Eckhoff et al.: Graph-based alert contextualisation, September 2025

The primary preprint groups alerts by relationships within time windows and uses Graph Matching Networks to relate groups to historical incidents. An alert graph is not automatically a full semantic ontology or provenance system. The source read is a preprint, and no peer-review status is inferred.

These papers give the project a clearer technical lineage than a generic “LLM for security” topic. The decisive practical questions are which logs/ground truth are available, who supports data engineering and secure environments, and how the student will separate research development from operational support. No project-specific compute allocation or graduate-outcome dataset was verified.

Funding, Application and Living Cost

The official job lists position code 1017 and annual gross salary NOK 550,800–595,000, depending on competence/experience, with 2% statutory pension contribution deducted. This is approximately NOK 45,900–49,583 gross per month by division by 12, before tax and other payroll treatment; it is not a take-home guarantee. The seed recruitment description lists full-time 37.5 hours/week. The official period is three years; a possible fourth year includes 25% other career-promoting work, such as teaching, supervision or research assistance, subject to qualifications and departmental needs. A second UiO PhD fellowship period is not available.

Housing materially affects buying power. Statistics Norway’s 2025 survey reports Oslo/Bærum monthly means of NOK 10,090 for a one-room dwelling and NOK 15,260 for two rooms. These are whole-dwelling averages, not shared-room quotes or 2027 offers; utilities vary across observed rents. They establish a useful scale, not an exact personal budget. Add food, transport, deposit, insurance and individual tax/payroll circumstances before estimating savings.

The SiO FAQ generally excludes PhD candidates from direct student-housing applications. UiO researcher housing guidance describes departmental routes but warns demand exceeds supply and contracts run 2–12 months. Plan private options from the start rather than rely on ordinary student housing for three years. The advertisement does not itemize tuition/registration charges, relocation, visa costs or conference/model budgets; request the actual doctoral-employment arrangements rather than import taught-degree fee assumptions.

Applicant Fit and Preparation

Actual qualifications and application documents

Required qualifications include a relevant Master’s/equivalent in cybersecurity, informatics, AI or a closely related field; foreign MSc-level education corresponding to at least four Norwegian-system years; strong CS or AI/ML foundations; strong programming/software development; and fluent English. Cybersecurity knowledge or demonstrated motivation to apply AI/ML to security is accepted in the wording. An unfinished Master’s must be completed by 31 December 2026. Normal academic thresholds are Bachelor’s average C, Master’s average B, and thesis B or better on the Norwegian scale; foreign applicants must explain their grading system.

LLMs/agents, semantic technologies, security operations, trustworthy AI and research experience are desired rather than all mandatory. A publication is an advantage, not an explicit minimum. A practical research/software portfolio could demonstrate log handling, evidence-linked decisions, reproducible comparisons and failure analysis; this is suggested preparation, not a new formal checklist.

Apply through the official recruitment system with motivation letter, CV, Bachelor’s/Master’s diplomas and transcripts, English evidence, works/software to be considered and 2–3 references. Documents must be in English or a Scandinavian language; the retained material does not specify translation or certification rules for documents originally issued in other languages. Approved English documentation is required before starting. The faculty currently lists TOEFL iBT 90, IELTS Academic 6.5, PTE Academic 62, or the listed Cambridge qualifications, with specific exemptions. An English-medium degree from China is not itself one of the stated exemptions; check the actual rule rather than assume a generic waiver. The job deadline is 30 November 2026; the seed specifies 23:00 Europe/Oslo, while the official PDF states the date, so use the live portal for the final cutoff.

Security eligibility and China-resident applicants

This is a real project condition: successful candidates must be security-eligible, background checks may be used, and suitability for cooperation with national-security stakeholders is assessed. The advertisement does not specify a clearance level or require Norwegian citizenship. It also does not provide a China-resident applicant's access decision.

NSM’s official guidance on foreign citizenship permits clearance following an individual overall assessment, taking account of the home state's security significance and personal links to that state and Norway; conditions may be considered. That is evidence against assuming a blanket foreign-national ban, and equally does not establish that an individual will qualify.

FFI’s clearance policy says work at FFI requires clearance/authorization and foreign-national cases can take longer, with difficult foreign records potentially affecting recruitment. It is dated 2019 and remains publicly available; it is general FFI policy, not proof that every UiO co-supervised student has the same employee/access arrangement. For this UiO position, ask recruitment which assessment and access are essential, whether a China-resident applicant can be considered, whether unclassified work can cover the whole thesis and how delays affect the latest start. Retain as a conditional opportunity; security feasibility for Chinese applicants remains unconfirmed pending project-specific clarification. No individual or nationality-group rejection is asserted.

Career Implications

ENISA’s skills guidance identifies wider cybersecurity workforce and competence gaps. This supports the relevance of security engineering, incident analysis and trustworthy AI skills, but is European context rather than a Norwegian vacancy count or a placement promise. Possible paths include security automation research, detection/response engineering, AI security evaluation and academic research. National-security jobs may retain access restrictions, so broad labour demand cannot establish access to all career paths. No verified programme salary or placement outcomes were obtained.

Next Actions and Questions to Verify

Before applying

Clarify the project's security/access requirements for a China-resident applicant, accepted English evidence, degree/grade equivalence and the live cutoff. Establish the latest start and the Master’s completion condition. Do not treat a general foreign-degree route as clearance.

At interview

Agree a tractable research question, accessible data/ground truth, compute/model budget, threat model and baseline comparisons. Clarify primary/co-supervisors, review frequency, human approval and safe experimental environments, plus open publication if sensitive access is restricted.

Before accepting

Obtain the salary placement, three- versus four-year workload, doctoral agreement, research/relocation costs, housing plan and immigration support. Establish what happens if security assessment delays access or the start. The opportunity has substantial research value, with a specific eligibility uncertainty that should remain visible in any public promotion.

Evidence and Sources

Recruitment, university and doctoral framework

• UiO official recruitment advertisement 309937 PDF
• Jobbnorge vacancy page 309937
• EURAXESS listing for recruitment reference 309937
• UiO Department of Informatics English overview
• UiO Digital Security institutional unit
• UiO Mathematics and Natural Sciences doctoral supplementary regulations
• ShanghaiRanking UiO profile read 17 October 2026

People, research and evaluation

• Vasileios Mavroeidis: UiO academic profile
• Gudmund Grov: personal academic homepage
• Experimenting with Neurosymbolic Artificial Intelligence for Defending Against Cyber Attacks
• Towards Agentic Investigation of Security Alerts
• A Graph-Based Approach to Alert Contextualisation in Security Operations Centres
• OASIS CACAO Security Playbooks 2.0 Committee Specification 01
• AgentDojo: dynamic prompt-injection evaluation for LLM agents
• Passing the Test You Trained On: Re-evaluating Prompt-Injection Detectors for LLM Agents

Eligibility, housing and careers

• NSM personnel-security guidance §8-7 foreign citizenship
• FFI official security-clearance explanation
• SiO doctoral housing eligibility FAQ
• Statistics Norway rental market survey 2025
• UiO accommodation for international researchers and staff
• ENISA cybersecurity skills and competences