Summary

• Capgemini, Wavestone, and France's national testing institution LNE will support the European Commission's AI Office.
• Their work will cover cyber threat modelling, evaluation tools, technical compliance methodologies, and assessments of advanced general-purpose AI.
• The assignment supports the AI Act's existing systemic-risk obligations, with regulatory authority remaining with the Commission.

The European Commission has appointed a consortium led by Capgemini to examine the cyber risks posed by advanced artificial intelligence models, adding specialist testing capacity as the EU begins enforcing the safety obligations placed on powerful general-purpose AI systems.

The French technology services company will work with consultancy Wavestone and France’s national testing institution, Laboratoire National de Métrologie et d’Essais (LNE). Their assignment covers cyber threat modelling, the development of evaluation tools, technical compliance methodologies, and assessments carried out at the request of the Commission’s AI Office.

Selected under Lot 2 of the Commission’s EC-CNECT-2025-OP-0032 procurement procedure, the consortium will concentrate on the potential use of advanced models for harmful cyber activity. Capgemini identifies automated vulnerability discovery, exploit generation, and the orchestration of attacks as capabilities requiring increasingly sophisticated assessment. The financial value of the contract and a detailed delivery timetable have not been disclosed.

The appointment comes as European regulators face the practical difficulty of assessing technology whose behaviour may change with successive model releases, additional tools, and different deployment conditions. The AI Act establishes legal responsibilities for model providers, but determining whether those responsibilities have been met requires technical evidence that can withstand regulatory scrutiny.

Evaluations move beyond documentation

General-purpose AI providers have been subject to obligations under the AI Act since August 2025, including requirements concerning technical documentation, copyright compliance, and information made available to downstream developers. Providers of models presenting systemic risks face additional duties covering risk assessment, mitigation, incident reporting, and cybersecurity safeguards.

The Commission’s enforcement powers over general-purpose models became applicable in August 2026. Those powers include requesting information, conducting model evaluations, requiring corrective measures, and imposing penalties where the relevant legal conditions are met.

Documentation can establish how a developer describes its testing procedures and risk controls, although it cannot independently demonstrate that a model behaves safely under every relevant condition. Technical assessments can provide a second source of evidence by examining what models are capable of doing and how reliably their safeguards operate.

Cyber capabilities are particularly difficult to evaluate because legitimate and harmful activities may involve similar technical operations. Identifying vulnerabilities can help defenders secure software, while the same capability could assist an attacker searching for weaknesses in systems they are not authorised to test.

A model’s ability to produce a plausible technical answer is also different from its ability to execute a sustained sequence of actions. Assessors must distinguish between generating instructions, completing individual technical tasks, and carrying out activities that require access to external systems or software tools.

Those distinctions influence the assessment of systemic risk. A model might demonstrate relatively limited capabilities in an ordinary conversational interface while becoming more effective when connected to automated testing tools or a system capable of executing code.

Regulators need methods that remain comparable

The consortium is expected to develop evaluation methods that the AI Office can use when examining emerging risks. Such methods must produce evidence useful for regulatory decisions while remaining adaptable to advances in model architecture and performance.

Testing is complicated by the variety of conditions under which an AI model can operate. A provider may impose access restrictions or monitoring controls, while an enterprise customer could integrate the same underlying model into a specialised application with different permissions and workflows.

The distinction between a model’s underlying capabilities and the additional risks introduced by an application affects regulatory responsibility. Providers of general-purpose models must address the obligations applying to their models, while organisations deploying AI systems retain responsibilities associated with their particular implementations.

Consistent testing procedures could help regulators compare evidence across providers and successive model versions. However, they will need to account for the possibility that a model’s performance changes after additional training, adjustments to its safeguards, or modifications to the tools available during evaluation.

The inclusion of LNE brings experience in measurement and conformity assessment to a consortium otherwise composed of commercial technology services businesses. That combination may help translate specialist cybersecurity testing into processes capable of producing repeatable evidence.

Wavestone contributes expertise in cyber risk and operational security, while Capgemini will coordinate the wider assignment. The individual testing procedures, reporting arrangements, and allocation of responsibilities between the participating organisations have not yet been made public.

AI safety becomes an operational requirement

The Commission’s wider regulatory approach already includes a Scientific Panel of independent experts and mechanisms for identifying systemic risks. The new consortium will provide additional practical assessment capability rather than replacing those advisory and supervisory structures.

In July, the Commission outlined an action plan addressing the relationship between advanced AI and cybersecurity. Its programme recognised that AI could strengthen defensive security operations while also enabling more capable attacks, creating a need for evaluation capacity that keeps pace with technical development.

For model developers, more systematic assessments may increase the importance of maintaining evidence about model behaviour, security testing, safeguards, and changes introduced between releases. Organisations building commercial applications on those models may also need greater clarity about their providers’ capabilities and limitations.

That does not mean the Commission’s technical evaluations will amount to a general certification of every AI application. Responsibility for enforcing the relevant rules remains with the designated authorities, while the consortium supplies assessments and methods to support their decisions.

The work also has implications for procurement. Businesses deploying advanced models may seek evidence that their suppliers can meet applicable requirements, but regulatory assessments alone cannot establish whether an individual deployment is secure. Application design, permissions, data handling, and operational controls remain separate considerations.

Capgemini, Wavestone, and LNE will now support the AI Office as it develops its technical approach to cyber-related systemic risks. The extent to which individual evaluation findings will become public, and how methods will evolve as model capabilities change, remain to be established.