Mistral AI has released the full weights of Mistral Large 4, its trillion-parameter flagship model, even as the Paris company says the model could help attackers carry out real cyberattacks, CSO Online reported Friday. Because downloaded weights cannot be recalled or updated, responsibility for the model's safeguards now shifts to every organization that runs its own copy.
What Happened
Mistral published the weights, the trained model's parameters that anyone can download and run, on Oct. 29, weeks after a public preview. Its modified MIT license permits use, modification, fine-tuning and commercial use; companies selling the model as a hosted service need a separate license.
In its release announcement, Mistral said monitoring of live preview traffic showed the model refusing black-box attacks against live targets, but that it still saw "a potential for uplift to conduct real attacks." It said it is changing its guardrails in response. Pierre Stock, Mistral's vice president of science, told CSO Online that later safeguard updates go to the company's managed API and future releases. Copies already downloaded will not receive them.
Mistral reports that Large 4 refused 96.6% of malicious cyber requests across three public test suites. It also scored 51% on the Artificial Analysis Cyber Index, which measures how well a model finds, reproduces and patches software flaws. CSO Online noted that Mistral's own chart shows two other open-weight models scoring higher on that index.
A version tuned for authorized penetration testing is in limited preview with vetted partners such as Orange Cyberdefense. Mistral said the European Union Agency for Cybersecurity (ENISA) is testing Large 4 to find vulnerabilities, audit code and analyze defensive logs.
Why It Matters
All of the benchmark figures come from Mistral, and analysts urged caution. Gartner principal analyst Niyati Daftary told CSO Online the refusal figure "tells you what their model does in their test scenario." Forrester's Charlie Dai said Mistral's coding claims are hard to validate without the technical report the company has promised but not dated.
For boards and executives, the practical issue is ownership. A company that self-hosts Large 4 takes on patching, monitoring and misuse prevention that a hosted provider would otherwise handle. Daftary advised running it as a controlled component, with isolated environments, least-privilege access and human sign-off, rather than as an autonomous agent.
The EU Question
The release is an early test of the EU AI Act. The European Commission's guidance on general-purpose AI obligations says models trained with more than 10 to the 25th power floating-point operations are presumed to carry systemic risk. Their providers must assess and mitigate those risks and report serious incidents. CSO Online noted that Mistral's post does not say whether Large 4 falls into that category. Stock said Mistral is a full signatory of the EU's general-purpose AI code of practice but did not address the classification.
The Primary Wire reported the preview Oct. 6, citing SiliconANGLE.
What to Watch
- Mistral's technical report, which Stock said will cover its cyber evaluations.
- Whether the European Commission treats Large 4 as a systemic-risk model.
- Results from ENISA's testing of the model.
