Europe’s AI sovereignty gap exposes a dependency on foreign models

At a glance

What this is: This is an opinion piece arguing that Europe is becoming dependent on foreign foundation models, with AI governance, public-interest obligations and infrastructure control all moving out of EU hands.

Why it matters: It matters to IAM, NHI and AI governance teams because model ownership, access paths, auditability and deployment control now shape who can govern AI systems and the data they touch.

Context

The core issue is not whether Europe can use AI, but whether it can govern the models, compute and data pipelines that increasingly mediate public information and automated decisions. When those layers sit outside the EU, strategic autonomy becomes a question of control over infrastructure, not just policy wording.

The article argues that regulation alone cannot offset dependence on foreign model ecosystems. For identity and AI governance teams, that means the control problem moves upstream into model lifecycle ownership, deployment boundaries and assurance over AI governance rather than only downstream compliance checks.

Key questions

Q: How should organisations govern AI systems that rely on external foundation models?

A: They should treat external foundation models as third-party critical infrastructure, not as ordinary software components. That means defining data residency, audit rights, change control, monitoring and recovery expectations before deployment. If the organisation cannot inspect or constrain the model lifecycle, it does not truly govern the system, even if it can consume the API.

Q: When does AI become a governance issue in cybersecurity operations?

A: AI becomes a governance issue when its outputs influence prioritisation, response, policy, or other decisions that security teams rely on. At that point, ownership, review, and accountability matter as much as accuracy. The programme must define who can trust the output and under what conditions.

Q: How do teams know whether AI governance is actually working?

A: Look for evidence that every AI interaction can be traced end to end, from identity and intent to output and enforcement. If auditors can ask for a transaction and receive a complete record in hours, not weeks, the programme is producing usable control evidence rather than just documentation.

Q: What should security teams do before adopting open small models?

A: They should verify the operating model, not just the parameter count. The key questions are who can deploy, update, fine-tune and revoke the model, where it runs, and what evidence exists for behaviour testing. Open models improve transparency, but they still need lifecycle control and access governance.

Technical breakdown

Why model dependency becomes a governance problem

Large language models are not just applications. They are infrastructure that shapes content, decisions and workflow outcomes at scale. When a small number of external providers control the model, training data, deployment options and update cadence, the organisation or jurisdiction consuming those models loses meaningful leverage over behaviour, resilience and auditability. That creates a structural mismatch between governance expectations and operational reality. In practice, the control plane is no longer just policy, it is ownership of model access, data residency and lifecycle decisions.

Practical implication: Treat model sourcing and deployment as governance assets, not just procurement choices.

Small open models change the control boundary

The article’s strategic alternative is not to clone frontier models, but to invest in smaller, open and inspectable models that can run on local or edge infrastructure. Technically, that changes the trust model. Smaller models can be embedded into environments with clearer data boundaries, tighter operational scope and more direct oversight of updates, fine-tuning and prompt handling. They are also more feasible for specific sectors and languages, which reduces the need to route sensitive content through remote black-box services. The trade-off is that governance must extend to versioning, evaluation and controlled distribution, not just access approval.

Practical implication: Build review and change-control around model versions, not only around who can call the API.

AI Act compliance needs lifecycle evidence

The piece is strongest when it links sovereignty to governance-by-design. Compliance for powerful models is not credible if it is bolted on after deployment. Practitioners need evidence of risk management, auditability and documented model behaviour throughout the lifecycle, from training data selection to deployment and monitoring. That is especially important where AI systems interact with public-interest content or personal data. Without that evidence trail, organisations may have policy statements but not demonstrable control. The governance gap is not the absence of rules, but the absence of lifecycle proof that the rules are actually enforceable.

Practical implication: Collect lifecycle evidence for model training, deployment and monitoring before the system is put into production.

Threat narrative

Attacker objective: The practical objective is not theft in the classic sense, but strategic control over the infrastructure that shapes information, automation and policy outcomes.

1. Entry occurs when public and private organisations adopt externally controlled foundation models as default infrastructure for content, analysis and automation.
2. Escalation follows when model providers, not the deploying organisation, effectively determine behaviour through weights, updates, data pipelines and opaque evaluation choices.
3. Impact emerges when dependence on foreign model ecosystems limits auditability, weakens sovereignty and leaves regulators writing rules for systems they do not control.

Breaches seen in the wild

• LiteLLM PyPI package breach: LiteLLM PyPI supply chain attack, credentials stolen from users.

Read and download The State of NHI & AI Agent Breach Report 2026, covering 200+ breaches impacting Non-Human Identities including AI Agents.

NHI Mgmt Group analysis

Strategic autonomy fails when the control point is externalised. Europe cannot credibly claim AI sovereignty if the core model stack, data pipelines and inference infrastructure sit outside its control. That is a governance problem before it is a market problem, because the organisation or jurisdiction consuming the model cannot fully inspect, constrain or recover it. Practitioners should read this as a warning that sovereignty is enforced through operating model design, not slogans.

Open small models are a governance strategy, not a cost compromise. The article is right to frame smaller, open and inspectable models as the realistic path for decentralised European environments. That matters because governance improves when model scope, update cadence and deployment location are easier to define. In identity terms, the question is who can authorise, attest and revoke model use across public and private workflows. Without that clarity, agentic AI identity risk becomes harder to contain.

Governance-by-design must become the default control philosophy. The article correctly argues that AI Act obligations cannot be satisfied by post-hoc documentation alone. Auditability, risk management and value alignment have to be built into the model lifecycle from the start, or compliance becomes performative. For practitioners, that means evidence of training data provenance, evaluation, monitoring and exception handling must exist before adoption, not after an incident or audit finding.

Europe’s AI problem is also an identity problem. Once models mediate what humans read, see and believe, the trust boundary shifts from the user interface to the system that decides what content is generated, filtered or amplified. That creates a non-human decision layer that must be governed as part of the identity estate, especially where model access, service accounts and delegated automation are involved. The practical conclusion is that AI sovereignty and NHI governance will converge in the same operating model.

What this signals

AI sovereignty will increasingly be judged by control over the model lifecycle, not by policy statements. Organisations that cannot evidence where models are trained, updated and deployed will struggle to prove governance in regulated environments. That makes procurement, architecture and audit design inseparable decisions.

Open and inspectable models create a narrower trust boundary, but they also move accountability closer to the deploying organisation. Once a team chooses to host or fine-tune models locally, it inherits responsibility for evaluation, monitoring and controlled change. The governance burden becomes more visible, not smaller.

Identity controls will matter more as AI systems start acting through delegated infrastructure. Service accounts, routing permissions and admin privileges will define who can alter model behaviour in production. That is where AI governance and NHI governance begin to overlap in practical terms.

For practitioners

• Define model sovereignty requirements Set explicit requirements for where model weights, training data, inference and logging may reside, and treat cross-border dependency as a governance decision rather than a technical convenience.
• Build lifecycle evidence into AI controls Require documented evidence for training data selection, evaluation, monitoring, exception handling and rollback before a model is approved for production use.
• Review who can operate model infrastructure Map the identities, service accounts and delegated admins that can change prompts, weights, routing or deployment settings, then limit them to the minimum operational scope.
• Prefer open and inspectable deployment patterns Use local or edge deployment options where sensitive content, regulated workflows or public-interest use cases need tighter oversight than remote black-box access can provide.

Key takeaways

• The article frames AI dependency as a governance failure, not just a market or technology choice, because external model control weakens oversight and recovery.
• Its central argument is that Europe needs inspectable, locally governable models if it wants compliance, public-interest alignment and operational autonomy to be credible.
• For practitioners, the takeaway is to govern model lifecycle, access and deployment boundaries together, because sovereignty claims collapse when those controls are split apart.

Related NHI Mgmt Group resources: AI Governance Guide · Agentic AI Identity Risk Board Briefing · Third-Party AI Risk Guide

Key terms

• AI sovereignty: AI sovereignty is the ability to control where a capability runs, who can use it, and who can revoke it. In practice, it depends on identity governance, jurisdictional awareness, and resilience controls that keep business workflows running when a provider changes the rules.
• Governance by Design: A control model that embeds standards, validation, and approval directly into the workflow where work is created or changed. Instead of relying on later review, it enforces acceptable state at the point of entry, which makes compliance and integrity part of normal operations.
• Model Lifecycle: The model lifecycle is the sequence of stages a model moves through, from rationale and development to validation, approval, deployment, monitoring, modification, and decommissioning. Effective governance ties controls to each stage so that the organisation can document purpose, measure performance, review risk, and preserve auditability over time.
• Foundation Model Dependency: Foundation model dependency is the condition where critical business or public-interest workflows rely on externally controlled models for inference, updates or content generation. It creates concentration risk because the consuming organisation may not control behaviour changes, data handling or recovery options.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, workload identity and agentic AI identity. It helps practitioners connect identity controls to the AI systems and service accounts their programmes increasingly depend on.