Two major AI companies have made cybersecurity central to their latest models, but are distributing them in very different ways. Google announced Gemini 4 Carbon on October 28 with access limited to selected defenders, while France’s Mistral AI released the downloadable weights of Mistral Large 4 on October 29.
Both models are meant to do more than explain security problems. They are built to help find weaknesses in software, test whether they can be exploited and develop fixes. The same abilities could be misused, so the key questions are who gets access, what safeguards apply and how the models are deployed.
Google’s Gemini 4 Carbon
Google DeepMind describes Carbon as its strongest model to date, with improved performance in coding, complex multi-step tasks and cybersecurity. It is built for long engineering jobs, such as tracing the root cause of bugs and developing fixes across large codebases.
Its security work runs in three stages: finding vulnerabilities, validating whether they can be exploited and producing patches.
Google reported scores of 88.2 percent on SEC-bench Pro, which tests finding vulnerabilities in real-world software, 54.8 percent on ExploitGym, which tests exploitation, and 71.0 percent on CWE-bench, which tests fixing weaknesses.
Google says Carbon outperformed Gemini 4 Argon, GPT-6 Astra and Claude Opus 5.5 on these benchmarks. These are results under particular test conditions, from Google’s own comparison.
Carbon is not available to everyone.
Access is initially limited to selected cybersecurity defenders and will widen gradually to trusted partners through the Fairwind Program, which Google introduced on September 2.
Fairwind is aimed at governments, critical infrastructure operators, healthcare providers, telecommunications services and technology platforms. It requires organisational vetting, user authentication and access controls, and bars sharing or reselling access.
Google lists safeguards including safety training, real-time defences against misuse, protection against prompt injection, monitoring of model behaviour and further testing before wider release.
It reported a 0 percent attack-success rate on the Gray Swan indirect prompt-injection benchmark, which applies only to the attacks tested and does not mean immunity. Its other safety results are company-reported.
Mistral Large 4
Mistral introduced Large 4, nicknamed Le Chonk, as a preview on October 6 and released its weights on October 29. It has about one trillion total parameters, around 49 billion active per token, a context window of up to one million tokens and text and image input. It uses a mixture-of-experts design, routing each token to four of 128 experts alongside a shared expert, so only part of the network works on each token.
Open weights let organisations download the model and run it on their own infrastructure. That gives them control over sensitive code and logs, room to customise it and independence from an outside provider. It is not fully open-source.
The model is released under the Mistral Ecosystem Open-Weight License, which carries conditions, and Hugging Face documentation puts the full model at roughly one terabyte of GPU memory, which is beyond most individuals and small organisations.
Mistral reported a 51 percent score on the Artificial Analysis Cyber Index, 82 percent on a task of reproducing a real open-source vulnerability and then patching it, and 93 percent on Cybench, which has 40 security challenges.
The company said internal tests found it useful for malware analysis, prioritising vulnerabilities, writing detection rules and supporting authorised penetration testing.
Safeguards and risk
Mistral reported an average refusal rate of 96.6 percent on malicious cyber requests drawn from JailbreakBench, StrongREJECT and AgentHarm, the highest it measured among open models. Test prompts do not guarantee refusal in real use. Mistral acknowledged that monitoring of live traffic showed the model could improve the ability to carry out real attacks, though it refused observed black-box attacks against live targets, and said it was modifying its safeguards.
It is also restricting some capabilities. Selected enterprise design partners, including Orange Cyberdefense, will get preview access to a specialised version for penetration testing, vulnerability assessment and red-team work. Access is vetted and contractual, and testing is limited to systems the organisation owns or has permission to test.
Two approaches
Google is managing risk by controlling who gets a powerful model, though that makes users dependent on the provider. Mistral gives organisations control, but the developer cannot control every later deployment.
Faster vulnerability discovery could help hospitals, banks, telecom firms and government systems find and fix weaknesses earlier, but attackers could try similar tools, raising the importance of timely patching. AI-generated fixes still need human review and testing, and organisations still need security staff.
