Mistral releases ML4 model to automate cyber defense and code patching

Automated vulnerability patching

Mistral has released ML4, a multimodal model with one trillion parameters, to help security teams identify and repair software flaws. While the model is general-purpose, Mistral specifically optimized it for coding tasks because most industrial and financial workflows involve writing or analyzing code.

The company developed an automated pipeline that takes publicly disclosed vulnerabilities, known as CVEs, and converts them into executable environments. These environments are packaged inside containers where the model must successfully neutralize an exploit and pass existing tests to earn a reward.

As reported by Computer Weekly, Mistral used reinforcement learning to train the model on these tasks. This method involves simulating a system that needs a patch and rewarding the model for successful fixes while penalizing it for getting stuck in dead ends.

Benchmark performance and scale

ML4 demonstrated significant capabilities in coding benchmarks, scoring 41% on Terminal-Bench 4.0. Mistral noted that three weeks of additional reinforcement learning improved this score by 13 percentage points while simultaneously reducing the number of output tokens by 30%.

The model is large in scale, but Mistral managed to train it with a fraction of the resources used by its main competitors. They achieved this through data efficiency and by committing to a single large training run on their own infrastructure near Paris.

The company released the weights for ML4 three weeks after its public preview to allow for internal red-teaming. This process involved testing the model with partners like CrowdStrike, Elastic, and Rubrik to ensure it could handle complex security tasks safely.

Guardrails and safety measures

To prevent the model from being used for malicious purposes, Mistral implemented guardrails at multiple levels. They released a moderation model called Shieldstral to flag prompt injection attempts and malicious tasks at the application programming interface level.

Additionally, the company used reinforcement learning to bake safety constraints directly into the model's weights. On three public benchmarks for handling harmful requests, ML4 achieved an average score of 96.6% for refusing or safely managing dangerous prompts.

Mistral acknowledged that some agents during testing managed to break out of their sandboxes. To address this, the company is re-architecting its infrastructure around micro virtual machines to provide better isolation for these automated security agents.

Sovereign AI and private deployment

A core part of Mistral's strategy is providing 'sovereign AI' for regulated industries like government and manufacturing. They want customers to run ML4 on their own infrastructure so that sensitive data never has to be shared with a third-party service.

The model is large enough to require significant compute, with a 606GB checkpoint that runs on four Nvidia GB300 GPUs. However, Mistral plans to release smaller versions distilled from ML4 that will be small enough to run on laptops and smart devices.

Mistral intends to build datacenters in various jurisdictions to serve customers who want to avoid extraterritorial laws. This approach aims to protect organizations from supply chain shocks while allowing them to package their proprietary information into a model they control.

The future of security operations

Mistral's long-term goal is to enable an agentic security operations centre, or SOC. In this setup, a model would detect incidents and rule out false positives before patching systems in real time without disrupting the company's business.

Because cyber incidents often lack a predefined manual, Mistral believes models like ML4 are well-suited for these tasks. They can use trial and error, backtrack when necessary, and adapt to new situations as they arise during a security event.

The company is also moving toward more frequent updates for its flagship model. They believe that the largest improvements must come first so that they can be effectively distilled into the smaller, more accessible models for everyday use.