Mistral Open-Sources Le Chonk While Warning of Attack Uplift

Mistral released Mistral Large 4's weights despite warning the model could uplift real cyberattacks, with future guardrail fixes limited to its API.

@sharedot

Mistral Open-Sources Le Chonk While Warning of Attack Uplift

By @sharedot · · 8 pages

  • AI
  • Open Weights
  • Mistral
  • Cybersecurity

Mistral released Mistral Large 4's weights despite warning the model could uplift real cyberattacks, with future guardrail fixes limited to its API.

What happened: weights out, warning attached

The trillion-parameter mixture-of-experts model is now downloadable, modifiable and self-hostable under a modified MIT license, with model-as-a-service providers needing a separate commercial license.

The company simultaneously conceded that its own preview monitoring saw "a potential for uplift to conduct real attacks," saying it would keep updating guardrails in the coming months — an unusual admission to pair with an open release.

The weights are available on Hugging Face and via Mistral's API.

The catch: guardrail fixes stop at the API

Mistral vice president of science Pierre Stock 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, organizations that download Le Chonk today will not receive future safety patches automatically, shifting responsibility for patching, monitoring and misuse prevention onto the operator.

CSO Online also notes the UK AI Security Institute found open-weight models trail closed frontier models on cyber tasks by four to seven months, creating what it calls "a persistent and irreversible risk of misuse" for highly capable open releases.

Evidence: strong refusal rates, contested claims

Mistral says Large 4 refused 96.6% of malicious prompts in the cyber-related portions of JailbreakBench, StrongREJECT and AgentHarm — the highest rate it has measured among open models — and scored 51% on the Artificial Analysis Cyber Index.

The company also claims leading agentic coding scores on Cognition's FrontierCode 2.0 and Terminal-Bench 4.0, though no technical report has yet been published.

Experts urged caution: Gartner's Niyati Daftary told CSO Online the 96.6% figure "tells you what their model does in their test scenario, not how it will behave with your data," while Forrester's Charlie Dai said the best-open-model claim remains hard to validate without the technical report.

The stakes: sovereignty vs. security

The release is a test of open-weight AI's appeal to enterprises and governments that want frontier cyber defense they can run on-prem and that cannot be cut off by a vendor.

Stock framed it as defensive empowerment: access to frontier intelligence that "can't be cut off by definition because the model is open-weighted."

But Forrester's Dai warned that self-hosting shifts responsibility for guardrails, access controls and misuse prevention to the operator, and that sovereignty and security "need to be evaluated together, not traded against each other."

Gartner's Daftary recommended running self-hosted deployments as controlled components with sandboxed code and human sign-off, not autonomous agents.

Questions the pack leaves open

Several claims are not yet verifiable. The ZDNET report puts ML4's Terminal-Bench 4.0 score at 41%, a one-point edge over GLM-5.3, while noting it trails GLM-5.3 by one point on the overall Coding Agent Index at 48 — and warns benchmark scores should be taken with a grain of salt.

CSO Online points out two open-weight models, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, score higher on the Artificial Analysis Cyber Index chart.

Stock told CSO Online a technical report covering cyber evaluation methods, refusal testing and preview monitoring is coming, but gave no date, and did not say whether the model is classified as systemic-risk under the EU AI Act.

What comes next

Adoption is already moving: ENISA, the EU's cybersecurity agency, said in a statement published with the announcement that it is testing the model for vulnerability identification, code auditing and defensive log analysis.

Mistral is keeping its specialized pen-testing and red-teaming variant behind vetting, with design partners such as Orange Cyberdefense getting preview access under contractual safeguards.

Meanwhile, the guardrail-versus-checkpoint split will keep defining the risk calculus for every organization weighing a self-hosted trillion-parameter model. Watch for the promised technical report and continued guardrail updates to Mistral's managed API in the coming months.

Sources

  1. techcrunch.com › Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity
  2. csoonline.com › Mistral open-sources AI model despite warning it could aid cyberattacks
  3. zdnet.com › Mistral's Le Chonk is now a top open model for coding, ranking alongside Codex GPT-6.1 Terra

More on AI