A cyberattack on France’s tax authority is pushing the government towards an unusually explicit form of technology sovereignty, with ministers preparing to use domestic artificial intelligence providers to probe weaknesses across public services rather than turning to the largest US AI companies.

French Budget Minister David Amiel said the government would hire “sovereign” AI providers, naming Mistral as an example, after the breach exposed taxpayer data and intensified scrutiny of ageing government infrastructure. Asked whether the programme could include OpenAI, Amiel said: “This excludes OpenAI.”

The response follows the theft of data relating to roughly 700,000 taxpayers from the French tax administration, while officials were also assessing a separate breach discovered this week. The first incident had already raised questions about how effectively vulnerabilities are identified across public systems containing some of the country’s most sensitive personal and financial information.

Although the immediate programme is a cybersecurity response, the procurement decision gives practical shape to a European debate that has often remained political. France has spent several years arguing that strategically important digital infrastructure should not depend entirely on overseas suppliers, but the harder test begins when that principle has to survive the technical and security requirements of operational government systems.

AI moves into the security stack

AI can support cyber teams by reviewing code and configuration data, identifying unusual patterns, mapping possible attack paths, and prioritising vulnerabilities for human investigation. Those uses do not remove the need for penetration testing, access controls, incident response, or specialist security staff, although they can increase the quantity of infrastructure that a limited team is able to examine.

That capability has particular value across public-sector estates built from a mixture of legacy applications, newer cloud services, specialist databases, and interfaces accumulated over decades. Amiel has separately described significant technical debt inside the French state, and placing a capable model over poorly documented or weakly maintained infrastructure will not remove those underlying weaknesses.

The choice of supplier therefore extends beyond model performance. A system examining government security weaknesses may require access to architecture, source code, logs, network information, or vulnerability data, making the location of processing, control of infrastructure, contractual access, and retention policies more consequential than they are in a routine office-assistant deployment.

Mistral has built much of its commercial proposition around giving organisations greater control over deployment, including private and self-hosted environments, while pursuing European enterprise and public-sector customers. France’s programme creates a prominent test of whether that proposition can translate from sovereignty language into practical security work across large and uneven government estates.

Sovereignty meets procurement reality

European governments have become more concerned about dependence on a small number of non-European cloud and AI providers as generative systems move into operational work. The concern covers data jurisdiction and resilience, but it also reaches bargaining power, supply continuity, technical lock-in, and whether public administrations retain enough expertise to understand the systems on which services depend.

France has now drawn a procurement boundary around a specific task rather than simply expressing a preference for European technology. Excluding OpenAI does not establish that a domestic provider will necessarily perform the work better, and sovereign sourcing does not make a system inherently secure, but it changes the criteria against which competing suppliers are judged.

Those criteria are likely to include where data is processed, who can access it, what happens to information submitted to a model, how activity can be audited, and whether the technology can operate without creating another external dependency. They also require European AI companies to demonstrate the less glamorous capabilities needed for government deployment — integration, security controls, predictable service, support, and the ability to work around existing systems rather than assuming a clean technical environment.

The tax-agency breach gives that debate an unforgiving setting because France is not buying sovereign AI simply to support a national champion. It is trying to find weaknesses after highly sensitive information was exposed, and the programme will be judged by whether vulnerabilities are discovered and fixed rather than by the nationality of the model provider.

If the tools accelerate remediation across the public estate, France could create a practical template for other European administrations trying to combine cybersecurity with technology sovereignty. If they merely add another software layer while technical debt remains untouched, control over the model will provide little protection against weaknesses elsewhere in the stack.