Summary
- France plans to use domestic AI providers, including Mistral, to test government services for cybersecurity weaknesses.
- The move follows theft of data concerning about 700,000 taxpayers from the French tax administration.
- A separate breach detected on 17 August was still being assessed when officials disclosed the new security programme.
France plans to use domestic artificial intelligence providers to test government services for cyber weaknesses after an attack on the tax administration exposed information concerning roughly 700,000 taxpayers.
Budget Minister David Amiel said the government would use what he described as sovereign AI providers and named Mistral as an example. He explicitly ruled out using OpenAI for the initiative.
The plan follows a breach of France’s tax administration disclosed last week. The Finance Ministry said data concerning about 700,000 taxpayers had been stolen, extending an incident already covered by Cyber Insider.
The investigation is also developing. Amélie Verdier, head of the French tax agency, said another data breach had been detected on Monday, 17 August, and remained under evaluation. Officials had not established publicly whether the two incidents were connected.
Sovereignty moves into security testing
Using AI systems to probe public services brings automated security testing into France’s wider policy of reducing strategic dependence on foreign technology providers.
The jurisdictional choice is particularly relevant to vulnerability research. Systems used to assess government infrastructure can be exposed to sensitive information about architecture, software behaviour, configuration weaknesses, and potential attack paths. The provider handling that work can therefore become part of the government’s own security dependency.
Domestic ownership does not by itself establish technical security. Deployment architecture, access controls, model boundaries, data handling, auditability, human oversight, and the process used to verify automated findings remain decisive.
The programme will also have to deal with the practical limitations of AI-assisted vulnerability discovery. Automated systems may accelerate reconnaissance and identification of suspicious behaviour, but findings still need to be reproduced, triaged, prioritised, and fixed. A high volume of low-quality findings can create its own burden, while poorly bounded testing can interfere with production environments.
Those concerns become more immediate when the initiative follows an actual compromise rather than a theoretical policy exercise. Tax administrations hold identity and financial information that can be valuable for fraud, social engineering, and intelligence collection, giving weaknesses in those environments consequences well beyond the availability of a government website.
The second incident disclosed this week also leaves conventional investigative questions unresolved. France has not publicly explained the initial access route, whether attackers retained persistence, whether the two breaches share infrastructure or methods, or who may have been responsible.
France is therefore changing part of its security-testing model before the full account of the incidents that prompted the response is known. The sovereign-AI decision addresses who will help search for weaknesses; the longer-term measure of the programme will be whether those weaknesses are removed and whether compromise can be detected and contained earlier.
