Mistral released the open weights of its trillion-parameter Large 4 model on Thursday, despite saying the model could give attackers "uplift to conduct real attacks." The Paris company published the download, which it nicknames "Le Chonk," weeks after a public preview during which it says it monitored cybersecurity-related use in live traffic. Anyone can now download, modify and run the model outside Mistral's control.
The core tension is that Mistral's safeguards do not travel with the file. Pierre Stock, the company's vice president of science, told CSO Online that the model ships with safeguards built into its weights, but that "subsequent safeguard updates are applied to our managed API and future model releases, rather than automatically updating downloaded checkpoints." In other words, any hardening Mistral does from now on reaches its hosted service, not the copies already on strangers' hard drives. The company says Large 4 refused 96.6 percent of malicious prompts in the cyber portions of three jailbreak benchmarks, the best rate it has recorded for an open model. Gartner analyst Niyati Daftary cautioned that the figure describes Mistral's own test, "not how it will behave with your data, your adversarial use cases or in production."
The release lands the same week that a separate demonstration showed what AI cyber capability now looks like in practice. Anthropic said on 27 October that one of its models carried out a successful remote attack against a Linux server, exploiting a kernel use-after-free flaw tracked as CVE-2026-98432. According to the company, the AI-generated code leaked a kernel address to defeat memory-layout randomisation, then overwrote a callback pointer to run commands as root. The address leak took 48 minutes and command execution five more, a total under an hour. Anthropic used an internal research model, not the public Claude, and ran against a server with a specific feature enabled, so the result does not generalise to every machine.
Taken together, the two events bracket the open-weights debate. In July the UK's AI Security Institute found that leading open-weight models trailed the closed frontier on cyber tasks by four to seven months and warned that for highly cyber-capable models, open release creates "a persistent and irreversible risk of misuse." Mistral is releasing an open model that it concedes could help attackers; Anthropic, which keeps its strongest models closed, is publicising how far such capability has come. Both companies frame their work as defensive. Mistral points to vulnerability-finding and patching as the main gains it observed, and is giving a specialised cyber variant only to vetted partners such as Orange Cyberdefense.
What remains open is verification and the law. Mistral has not published the technical report behind its "best open-weight model for coding" claim, which Forrester's Charlie Dai said leaves it "difficult to validate across workloads." Under the EU AI Act, models trained above 10^25 floating-point operations are presumed to carry systemic risk, with duties to assess and mitigate it; Mistral's post does not say whether Large 4 crosses that line, and Stock would not say whether it has been classified as systemic-risk, noting only that the company is a full signatory of the EU's general-purpose AI code of practice. The weights, meanwhile, are already downloaded. Whatever is decided about them applies to copies no one can recall.