Skip to content
IIAIT

AI Today · 30 October 2026

Mistral publishes the weights of Large 4 and researchers expose flaws in AI assistants' built-in browsers

Today: a European trillion-parameter model that organisations can run on their own servers, security weaknesses in the browsers built into five AI coding assistants, and a CrowdStrike report on an espionage campaign driven by AI agents that also reached European governments.

  1. 01

    Mistral releases open weights for Mistral Large 4, a one-trillion-parameter model

    French company Mistral AI released the weights of Mistral Large 4 on 29 October, about three weeks after a preview. It is a mixture-of-experts model with about one trillion parameters, of which around 52 billion are active per step, accepts text and images, and was trained on Mistral's European infrastructure. The weights are on Hugging Face under Mistral's Ecosystem Open-Weight licence, which allows commercial use but requires a separate licence for companies offering the model as a hosted service. Through Mistral's API it costs 0.70 dollars per million input tokens and 2.10 dollars per million output tokens; the coding and security benchmark results Mistral cites are its own.

    What it means for you

    If data must stay on your own infrastructure or within the EU, this is a model you can host yourself, although a model of this size needs substantial GPU capacity. Read the licence terms before building a service on it, and test it on your own tasks rather than relying on the vendor's benchmarks.

    Sources

  2. 02

    Researchers show that the built-in browsers of five AI coding assistants can expose files on a developer's computer

    Forever Security published research on 29 October showing weaknesses in the browsers embedded in Cursor, GitHub Copilot, OpenAI Codex, Visual Studio Code and Claude Desktop. In its tests, a web page or a local file in a project could run a command, show a fake address in the address bar, or read other files on the disk, such as SSH keys and .env files. According to the researchers, Cursor and GitHub have fixed their issues, Codex and VS Code were still unpatched on 29 October, and Anthropic accepted a related report from another security firm, with no fix confirmed in the write-up.

    What it means for you

    Open projects from unknown sources in these tools with care, and be cautious about clicking links that an assistant suggests. Keep the tools updated, and do not keep passwords, keys or tokens in plain project files where an exposed browser could read them.

    Sources

  3. 03

    CrowdStrike: China-linked group used AI agents against more than 200 organisations, including European governments

    CrowdStrike reported on 29 October that VAPOR PANDA, a group it links to China, used an AI agent framework called ARTEX against more than 200 organisations in at least 14 countries within 11 days from 17 October. The agents, mainly running on the DeepSeek model, handled reconnaissance and movement inside networks while a human operator steered them at key points. CrowdStrike confirmed activity after a break-in at nine organisations; targets included foreign ministries and other government bodies in Europe. Many of the entry points were old weaknesses, such as leaked VPN keys and stale credentials from earlier data exports.

    What it means for you

    AI agents let attackers try known weaknesses against many organisations at once, so basic hygiene matters more. Check that keys and passwords exposed in past leaks or old backups have been changed, and that internet-facing servers and devices are patched.

    Sources