Mistral AI is working on deploying an AI model to European banks to detect security vulnerabilities in the code of their IT infrastructures, according to an indiscretion from Bloomberg. It is impossible to say when this model will be launched to attack vulnerabilities, nor since when these discussions began with the establishments.
« Imagine-t-on les bases de données de l’armée française scannées par Mythos ? »
But what is certain is that Mistral did not wait for the presentation of Mythos at the beginning of April to tackle this issue of cybersecurity. The startup was already working with its clients in the banking sector to discover vulnerabilities in the code of their software before the first steps of the Anthropic model, according to our colleagues. From now on, the aim is to develop a “turnkey” version for wider deployment.
« We have one of our competitors who knows how to do fear marketing very well.“, testified (without mentioning the name of Anthropic) Arthur Mensch, the general director of Mistral, before the commission of inquiry into digital vulnerabilities at the National Assembly this week. “We work with our clients to help them on these cyber issues“, he adds. For him, it is a regal subject “and an additional argument to say that we must have control over this technology » :
« You cannot have the French army databases and code scanned by Mythos. This creates such an irremediable dependence that solutions absolutely must be found. »
He could have added: sovereign solutions, because that is of course what we implicitly understand. Mythos is currently only available in a preview format and distributed sparsely to predominantly American organizations and companies. And Europe still hasn’t received its club card.
An army of cybersecurity officers at Microsoft
After Mythos and OpenAI’s Daybreak initiative, and before that of Mistral, Microsoft has lifted the veil on its new multi-model agentic security system. The editor is not content with just one model: several are at work and they are supported by an army of more than a hundred agents specialized in attacking the code.
MDASH, for “Microsoft Security multi-model agentic scanning harness”, has already found 16 vulnerabilities in the authentication stack and network infrastructure of Windows, including 4 critical flaws allowing remote code execution. “The strategic implication is clear: the discovery of vulnerabilities by AI has moved from the stage of laboratory curiosity to that of a defense tool deployable on a large scale in companies », explained l’éditeur, qui vante «Âthe lasting advantage » of the son système agentique.
The system is based on three elements: a set of complementary AI models driven by MDASH, including several specialized models which confront each other to identify vulnerabilities; the hundred agents dedicated to specific tasks (detection, verification, potential exploitation of a bug); and finally an analysis chain that can be extended with plugins.
On the CyberGym benchmark, which brings together more than 1,500 tasks reproducing real vulnerabilities from open source projects, MDASH achieves a success rate of 88.45%. This is the best score published to date, and about 5 points more than Claude Mythos Preview, and 6 more than GPT-5.5, its closest competitors.
Microsoft emphasizes that this result was obtained with AI models already publicly available, which suggests that the performance comes mainly from the “agentic” orchestration around the models rather than the models themselves.
MDASH agents are used by Microsoft security engineering teams and tested by a group of customers in a private preview. It is possible to request access to the preview.
/2025/09/05/2-laure-avatars-live-68bac2d061625024390257.png "DIRECT. War in the Middle East: China calls for a lasting ceasefire and the reopening of the Strait of Hormuz \"as soon as possible\""){kind=avatar}
