Too Smart To Help You Defend

On 6 October Mistral, the French AI lab, released a public preview of Mistral Large 4, a model it says has one trillion total parameters with about 49 billion active, trained from scratch on 3,800 NVIDIA Grace Blackwell chips in its own European datacenters. Mistral plans to release the weights by the end of October; Reuters puts the date at 27 October. Independent benchmarker Artificial Analysis scored it 82 percent on a test that asks a model to reproduce a real flaw in open-source software and then patch it. Mistral says several leading closed models score near zero on that same test because they refuse the task.

A safety filter is a decision about who gets help, and most companies never made that decision. Their vendor made it for them. Stephen shares his own frustration on company and client work: mention cyber, and the model shuts the door, even on purely defensive tasks.

In this episode, Stephen Forte covers:

• What Mistral released. A very large model, weights coming for self-hosting, and early access with reduced moderation for cybersecurity leaders, vetted partners and state authorities.
• The one independent score. 82 percent on Artificial Analysis's reproduce-and-patch test and 50 on its broader cyber index, strong but not first. Mistral's own long list of scores is left aside.
• The refusal claim. Mistral says closed models including Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse. That list comes from Mistral, not Artificial Analysis.
• Losing the tool mid-incident. Mistral's underrated point: an assistant that goes silent during a crisis is itself a security risk.
• My own wall. Pen-test summaries, fix write-ups and scan reports for the board all mention vulnerabilities, and all can trip the filter. The block is a club, not a scalpel.
• The locksmith problem. A model smart enough to find zero-day flaws should be smart enough to tell defending your own company from attacking someone else's.
• The opening for open models. When you hold the weights, you set the policy. Watch for security teams quietly working around refusals with tools nobody approved.
• Three questions for your security chief. Which tasks does our vendor refuse? Do we have a self-hosted path under our own policies? Who signs off on access with fewer restrictions?
• The takeaway. The safety filter that blocks the attacker also blocks the defender. Someone on your payroll has to own that tradeoff.

Sources:

• Mistral AI, "Introducing Mistral Large 4," 6 October 2026: https://mistral.ai/news/mistral-large-4/
• Artificial Analysis, Mistral Large 4 analysis, 6 October 2026: https://artificialanalysis.ai/articles/mistral-large-4-france-ai
• Reuters via CNA, "France's Mistral launches AI model it says outperforms some Chinese rivals," 6 October 2026: https://www.channelnewsasia.com/business/frances-mistral-launches-ai-model-it-says-outperforms-some-chinese-rivals-6435786

The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.