Senior AI Engineer – Cybersecurity

Spektrum supports apex purchasers (NATO, UN, EU, and National Government and Defence) and their Tier 1 supplier ecosystem with a wide range of specialist services. We provide our clients with professional services, specialised aerospace and defence sales, delivery, and operational subject matter expertise. We are looking for personnel to join our team and support key client projects.

Who we are supporting 

The NATO Communication and Information Agency (NCIA) is responsible for providing secure and effective communications and information technology (IT) services to NATO's member countries and its partners. The agency was established in 2012 and is headquartered in Brussels, Belgium.

The NCIA provides a wide range of services, including:

• Cyber Security: The NCIA provides advanced cybersecurity solutions to protect NATO's communication networks and information systems against cyber threats.
• Command and Control Systems: The NCIA develops and maintains the systems used by NATO's military commanders to plan and execute operations.
• Satellite Communications: The NCIA provides satellite communications services to enable secure and reliable communications between NATO forces.
• Electronic Warfare: The NCIA provides electronic warfare services to support NATO's mission to detect, deny, and defeat threats to its communication networks.
• Information Management: The NCIA manages NATO's information technology infrastructure, including its databases, applications, and servers.

Overall, the NCIA plays a critical role in ensuring the security and effectiveness of NATO's communication and information technology capabilities.

The program

Assistance and Advisory Service (AAS)

The NATO Communications and Information Agency (NCI Agency) is NATO’s principal C3 capability deliverer and CIS service provider. It provides, maintains and defends the NATO enterprise-wide information technology infrastructure to enable Allies to consult together under Article IV, and, when required, stand together in the face of attack under Article V.

To provide these critical services, in the modern evolving dynamic environment the NCI Agency needs to build and maintain high performance-engaged workforce. The NCI Agency workforce strategically consists of three major categorise's: NATO International Civilians (NIC)'s, Military (Mil), and Interim Workforce Consultants (IWC)'s. The IWCs are a critical part of the overall NCI Agency workforce and make up approximately 15 percent of the total workforce.

Role ID – 2026-0137

Role Background

The NATO Cyber Security Centre (NCSC) is a team of over 200 members working to monitor and protect NATO networks. In the NCSC’s role to deliver robust security services to the NATO Enterprise and NATO Allied Operations and Missions (AOM), the centre executes a portfolio of programmes and projects around 219 MEUR euros per year, in order to uplift and enhance critical cyber security services.

The Portfolio ranges from Programme of Work (POW) activities funded via the NATO Military Budget (MB) to Critical / Urgent Requirements (CURs/URs) and NATO Security Investment Programme (NSIP) projects funded via the Investment Budget(IB). In some edge cases, projects are also funded via the Civilian Budget (CB). Projects can span multiple years and are governed by various frameworks, including the Common Funded Capability Development Governance Framework (CFCDGM).

In order to execute this work, the NCI Agency requires support with the work undertaken by the NATO Cyber Security Centre (NCSC) in the area of Communications and Information System (CIS) security, cyber defence and cyberspace operations. This Statement of Work (SoW) specifies the required skillset and experience.

Objectives

This Statement of Work (SoW) outlines the services to be provided by the Supplier to enable and operate an on-premise AI server capability supporting cyber security services, including SOC, Digital Forensics, Malware Analysis, Threat Hunting and Incident Response.

The work focuses on:

• Implementing and optimizing AI use cases aligned with cyber security operations,
• Proposing and testing new models and approaches
• Integrating relevant data sources
• Designing and implementing a robust Retrieval-Augmented Generation (RAG) pipeline,
• implementing authentication and limitation of rights (role-based access, least privilege), and performing a risk analysis on Confidentiality, Integrity and Availability (CIA) for the data, toolset and models used.

They shall integrate with existing operational processes and documentation (e.g., existing SOPs/SOIs in Confluence, existing access management processes, logging/monitoring practices) and shall prioritise updating/aligning existing documentation rather than creating parallel artefacts

Deliverables

Deliverable 1:

Capture the current state (“as-is”) of the on-premise AI server capability and the supporting processes, including:

• Current architecture and deployment (hardware, virtualization/containers, GPU stack, storage, network zoning, backups, patching approach).
• Current AI toolset (frameworks, model runtimes/serving, vector DB, embedding models, LLMs, orchestration, prompt tooling, pipelines).
• Current security controls (authentication, RBAC, secrets management, certificate management, host hardening, network controls).
• Current data sources used or planned for use (SOC telemetry, EDR, SIEM, ticketing/case mgmt, threat intel, forensics repositories, malware sandboxes, knowledge bases/Confluence, etc.).
• Current logging/monitoring (system, application, model usage/audit logs), incident handling integration.
• Current documentation: review and map existing SOPs/SOIs in Confluence and identify documentation gaps and misalignments with actual practice.
• Workshops: organise up to 3 workshops with stakeholders (SOC/DFIR/Threat Hunting/IR/Platform admins) to validate the as-is workflow and priorities.

Deliverable 2:

Based on the approved recommendations from D1, implement agreed improvements to make the AI server capability operational and secure, including:

• System hardening and baseline configuration (OS/container runtime, GPU drivers, patching approach).
• Authentication integration (e.g., enterprise IdP/LDAP/AD as applicable).
• Limitation of rights: RBAC roles, least privilege, separation of duties (admins vs users vs auditors) and how this applies to specific AI dataset.
• Secrets management and secure configuration handling.
• Implementation of a multi-model multi user approach to best serve the potential use cases.
• Audit logging enabling traceability of:
• user access,
• data access,
• model usage (prompt/response metadata as policy allows),
• administrative actions.
• Update existing SOPs/SOIs in Confluence to reflect the implemented operational model (do not create parallel SOPs unless required).

Deliverable 3:

Design and implement (as agreed) data integrations and an operational RAG pipeline to enable and optimize AI use cases, including:

• Identify priority use cases and required data (e.g., alert summarization, case enrichment, IOC/TTP retrieval, playbook guidance, forensic artefact Q&A, malware triage support, threat hunting hypothesis support).
• Implement data ingestion/connectors (as feasible) aligned with existing systems and permissions such as SIEM, Forensics Lab, internal wiki etc.
• Build RAG pipeline components:
• ingestion, chunking strategy, metadata schema,
• embedding strategy and vector store configuration,
• retrieval strategy (filters, hybrid search where applicable),
• grounding/citation strategy,
• evaluation approach (quality, hallucination reduction, regression tests on curated datasets).
• Propose/validate models:
• recommend model families/serving approach suitable for on-prem constraints,
• propose new models where beneficial (e.g., embeddings, rerankers, small task models),
• optimize existing ones (latency, throughput, context management).
• Update existing documentation (SOPs/SOIs) to include operational steps for maintaining the RAG pipeline and integrations.

Deliverable 4:

Conduct a risk analysis focused on Confidentiality, Integrity and Availability (CIA) for the AI server environment, considering:

• Data classification and sensitivity (SOC data, IR case data, forensics artefacts, malware samples, intel feeds, internal knowledge bases).
• Threat model relevant to on-prem AI workloads (insider misuse, prompt injection/data exfiltration, supply-chain, model poisoning, insecure connectors, misconfiguration, credential theft, lateral movement).
• CIA risks introduced by:
• selected models and runtimes,
• RAG pipeline and vector store,
• ingestion pipelines and connectors,
• storage and backup strategy,
• authentication/RBAC design and audit logging,
• admin operations and patching.
• Mitigation plan: technical, procedural and documentation changes, aligned with existing security governance.

Deliverable 5 (Training package and knowledge Transfer):

Best usage of the on-premise AI server

• How to access the platform (authentication, MFA where applicable) and comply with limitation of rights (RBAC).
• Operational do’s/don’ts: data handling, prompt hygiene, safe usage patterns, auditability expectations.
• How to use and interpret logs/audit trails relevant to AI usage (as permitted by policy).
• How to report issues, request access changes, and request new integrations/models through existing processes.

How to extend existing AI use cases

• How to onboard a new use case (intake template, success criteria, risk considerations).
• How to add/modify RAG sources (connectors, metadata, access filters, retention).
• How to adjust chunking, embeddings, retrieval strategies, and citation/grounding.
• How to introduce evaluation and regression tests for use cases (quality and safety).

Fine-tuning adaptation fundamentals

• When fine-tuning is appropriate vs RAG / prompt engineering / tool use.
• Data preparation and governance (dataset curation, PII/secret handling, licensing).
• Fine-tuning workflow on-prem (as supported by the implemented toolset): training runs, validation, versioning, rollback.
• Model lifecycle management: documenting changes, performance tracking, and approval workflow.

Use of commercial off-the-shelf products for Cyber Threat Intelligence with non-classified data

• Practical usage patterns for tools such as Anthropic Claude (and equivalent services) strictly with non-classified data.
• Data handling rules and redaction/anonymisation approach (what must never be submitted).
• Example CTI workflows (e.g., summarising public reports, extracting IOCs from open sources, mapping to ATT&CK, drafting non-classified briefs).
• Decision guidance: when to use on-prem models vs external COTS services.

Training delivery requirements

• On-site delivery.
• Approx. 20 participants.
• Interactive format including demonstrations and hands-on exercises using approved datasets and non-sensitive content.

Deliverable 6:

This part of the work is dependent on work already done under D2, D3 and D4 and identified as agreed follow-up activities in the realm of:

• implementation support
• tuning
• maintenance
• monitoring improvements
• connector/model updates
• documentation updates

Essential Skills, Experience and Certifications

• Demonstrable 3+ years experience as a Red Hat Linux system administrator / platform engineer operating on-prem environments (servers, storage, networking, hardening, patching, backup/restore).
• Demonstrable 1+ years experience building and operating AI/ML platforms on-prem (GPU servers, CUDA stack, containers, model serving).
• Demonstrable 2+ years experience integrating enterprise authentication and authorization (AD/LDAP/SAML/OIDC), including RBAC and least-privilege design.
• Demonstrable experience delivering AI-enabled cyber security use cases in at least two of the following domains:
• SOC
• Threat Hunting
• Incident Response
• Digital Forensics
• Malware Analysis
• Demonstrable experience implementing Retrieval-Augmented Generation (RAG) in production-like environments, including:
• ingestion pipelines, chunking/metadata design,
• embeddings/vector databases,
• retrieval tuning (filters/hybrid search/reranking),
• grounding/citations and evaluation methods.
• Demonstrable experience integrating and governing multiple data sources, such as:
• SIEM/SOAR,
• EDR,
• case management,
• threat intel,
• knowledge bases (e.g., Confluence),
• file shares/object stores,
• forensic repositories,
• while enforcing access controls.
• Demonstrable experience with security logging and auditability for AI systems (access logs, admin activity logs, model usage telemetry as permitted by policy).
• Demonstrable experience producing and maintaining operational documentation in Atlassian Confluence (minimum 2 years), including SOP/SOI style documentation and runbooks.
• Good knowledge of OSI layers and core protocols (TCP/IP, VLANs, routing basics, TLS).
• Strong knowledge of containerization (Docker/Podman) including GPU scheduling concepts.
• Strong knowledge of model serving patterns (e.g., vLLM/TGI/llama.cpp-class runtimes, API gateways/reverse proxies, rate limiting).
• Knowledge of LLM security risks and mitigations:
• prompt injection,
• data exfiltration via RAG,
• insecure connectors,
• poisoning (data/model),
• supply-chain risks,
• secure prompt/data handling and redaction practices.
• Ability to perform a structured CIA risk analysis (Confidentiality, Integrity, Availability) for the chosen toolset/models/data flows, and translate this into actionable mitigations.
• At least one relevant certification in security / cloud / platform / AI security, such as:
• CISSP, CISM, CCSP
• GIAC (e.g., GSEC, GCED, GCIH, GMON, GCIA, GDSA, etc.)
• Linux: RHCSA/RHCE or equivalent
• Equivalent certifications may be accepted if demonstrably relevant.

Working Location

• Mons, Belgium

Working Policy

• Hybrid (50% remote, 50% onsite)

Travel

• Frequent travel to NATO site in Mons is required

Security Clearance

• Valid National or NATO Secret personal security clearance