A hacker using the handle mrwho posted an offer on a cybercriminal forum on September 16, 2026, claiming to sell what they describe as the complete source code repository of Mistral AI, the French artificial intelligence company that competes with OpenAI and Google DeepMind.
The listing, titled "Selling mistral.ai Source Code," appeared on a forum where stolen data and access credentials are regularly traded. FrenchBreaches reviewed the seller's directory tree and a code sample to assess the claim.
What the 339-File Directory Contains
The attacker shared a file listing of 339 items, mostly archives corresponding to internal development repositories. Many carry names that explicitly mark them as non-public: mistral-inference-private, mistral-inference-internal, mistral-finetune-internal, mistral-common-internal, vllm-internal, and client-python-private.
Other entries point to infrastructure and operations tooling, including sre-infra, sre-kubernetes-tools, network, model-delivery, registry, platform-docs, and several repositories tied to model training and inference. A handful reference specific industries or functions, such as finance, KYC, legal, billing, cybersecurity, and customer service, though the directory names alone do not prove those repositories contain client data.
The Webstral Sample: A Working Browser Agent Prototype
FrenchBreaches examined a sample file named webstral. It contains functional code for a browser-based AI assistant built as a Chrome extension. The tool lets a user instruct an AI model to perform actions on the web: navigate to a URL, read page content, click buttons, and fill out forms.
The code calls Mistral's API directly and references the devstral-medium-2507 model. An API key must be supplied by the user and is stored locally in the browser. No active Mistral API keys were embedded in the reviewed sample.
Webstral stands out because it is not a public library. It implements a complete agent loop: page analysis, automatic action suggestions, conversation history management, and direct interaction with DOM elements. The prototype can search for information, compare services, and complete forms based on user-provided context. This suggests the leaked archive may contain internal experiments and development tools that were never intended for public release.
A Second Breach, or the Same One Recirculated?
The seller claims Mistral was compromised first in May and again in September. A previous incident did hit Mistral's development environment earlier in 2026. The samples reviewed by FrenchBreaches confirm that someone possesses code apparently sourced from Mistral's internal repositories, but they do not independently prove a fresh intrusion occurred in September.
One possibility is that data stolen in the May incident is being resold under the guise of a new breach. Finding files, commits, or modifications dated after the May compromise would be the strongest evidence of a second intrusion. Without that, the September claim remains unverified.
No Evidence of Customer Data Exposure
So far, the examined samples cover source code and development projects. No large database of user personal information was identified. The company names appearing in some project titles do not, by themselves, confirm that those organizations' data was compromised.
The primary risk sits in intellectual property and technical exposure: internal architecture, prototype logic, development workflows, and potentially sensitive configuration details. For a company like Mistral AI, whose value depends heavily on model training techniques, inference infrastructure, and proprietary tooling, that kind of leak can be damaging even without customer records.
What Teams Should Watch For
If the leak is authentic, the exposed repositories could reveal how Mistral structures its fine-tuning pipelines, manages GPU clusters, and deploys models into production. Internal Kubernetes tooling, network configurations, and registry infrastructure would give competitors or attackers a detailed map of the company's operational stack.
The prototype browser agent code also raises questions about Mistral's product roadmap. Webstral appears to be a fully functional agent framework, not a research demo. Its existence suggests Mistral was exploring direct browser integration for AI-powered task automation, a direction that overlaps with what Anthropic and OpenAI have been building with their own computer-use and browsing tools.
Mistral AI has not publicly commented on the claim as of publication. The company did not respond to a request for comment.