- Google is limiting Gemini 4 Argon while it tests cyber safeguards.
- Gemini 4 Argon replaces Gemini 3.5 Pro and rivals OpenAI and Anthropic models.
Google has introduced Gemini 4 Argon, its latest frontier AI model, but is initially restricting access to a group of cybersecurity organisations while it continues testing safeguards ahead of a broader release.
Unlike a conventional Gemini launch, Argon is not immediately available to developers, enterprises, or consumers. Google said models with capabilities at this level require a phased release, with selected cyber defenders receiving access first through its Fairwind Program.
The company has not provided a date for general availability. It plans to expand access first to paid API customers and Google AI Ultra subscribers before making Argon more widely available.
Google is also participating in a US government process that gives officials pre-release access to advanced AI models. The Argon launch came one day after Google and several other AI companies joined a voluntary US safety agreement covering pre-release evaluations and safeguards against unintended hacking or unauthorised system access.
Why Google is restricting access to Argon
Google said it is continuing to strengthen safeguards against several categories of misuse before Argon becomes broadly available. These include cyberattacks and chemical, biological, radiological, and nuclear-related risks.
The company has tested those protections through internal and external red-team exercises that attempt to find ways around a model’s safety controls. Google said the testing combines manual attacks with automated methods designed to identify weaknesses before wider deployment.
Argon also includes protections against indirect prompt injection, where instructions embedded in external content attempt to redirect an AI system away from the task requested by the user.
Its defences against indirect prompt injection include adversarial training and automated red-team testing, according to Google.
Another safeguard monitors Argon’s reasoning and actions while it is carrying out a task. Google said the system is intended to stop execution when the model begins operating outside the user’s intended objective.
The company is also using isolated environments for higher-risk training and evaluation. Google said these environments are intended to contain advanced models while safety testing takes place.
Feedback from early deployments will also be used to refine Argon’s safeguards before access expands, Google said.
Google trained Argon for cybersecurity tasks including finding, validating, and patching software vulnerabilities.
Its initial rollout builds on Fairwind, a limited-access programme Google launched on September 2 for governments, Google Cloud customers, and selected cybersecurity partners. Google said the programme is intended to give defenders an “adaptation window” to strengthen systems before malicious actors gain access to comparable capabilities.
The programme initially provided access to Gemini 3.8 Flash Cyber and Google’s CodeMender system, which can identify vulnerabilities and generate and validate patches.
Google said Fairwind had more than 650 participating partners globally when the programme launched. Initial participants span government and national cybersecurity authorities, critical infrastructure operators, technology companies, and security providers.
Access is subject to operational restrictions. Participating organisations must limit use to employees working in areas such as cybersecurity, incident response, or penetration testing, while controls including multi-factor authentication are required.
Argon introduces another distinction within that controlled-access model. Google said trusted Fairwind users and its own internal security teams will receive Argon without cyber guardrails, allowing them to use its full cybersecurity capabilities for defensive work.
The company is continuing to strengthen cyber safeguards before making Argon broadly available.
Cloud security company Wiz is among the organisations testing Argon through its Scan for Good programme.
According to Google, Argon identified a vulnerability involving sensitive personal information in healthcare software used by hospitals after earlier frontier models had failed to detect it. Google did not identify the other models involved in that comparison.
Google reported that Argon scored 68% on CWE-bench v1, a benchmark that evaluates whether AI systems can remediate software vulnerabilities.
The company also tested the model against an internal benchmark covering codebases written in 20 programming languages. Because both results come from Google’s own testing, they should be treated as company-reported performance rather than independent assessments.
Google said Argon can autonomously find, validate, and patch critical software vulnerabilities.
Its broader release will include cyber safeguards intended to restrict harmful use of those capabilities, according to Google.
Gemini 4 follows a delayed model roadmap
Argon’s arrival also changes Google’s previously announced Gemini roadmap. The company has scrapped Gemini 3.5 Pro, a model that CEO Sundar Pichai had earlier said was expected in June.
Google instead spent subsequent months introducing smaller Gemini Flash models before moving directly to Gemini 4 Argon at the top end of the family. Ars Technica reported that Gemini 3.5 Pro never materialised despite Google’s earlier plans for the model.
Reuters reported that Gemini 4 followed months of delays. During that period, Google overhauled DeepMind, while several leaders involved with Gemini left the company. Google has not said that the organisational changes or Argon’s safety testing caused the delays.
The cancellation of Gemini 3.5 Pro leaves Argon as Google’s next highest-end model after several months in which new releases were concentrated around smaller Flash variants.
Outside cybersecurity, Google is positioning Argon for software engineering and professional work, including legal and financial tasks.
The model increases Gemini’s maximum output from 64,000 tokens to one million tokens. That higher limit is designed to support longer outputs and extended tasks without reaching the previous ceiling.
Google reported a score of 77.9% on DeepSWE v1.1, which tests AI models on longer software engineering tasks. Google’s published benchmark comparisons place Argon against frontier models from OpenAI and Anthropic, with results varying by test.
Google is already using the model internally. The company said Argon analysed fleet-wide data-centre telemetry and identified memory optimisations that freed more than 300 TiB of memory after deployment.
Its engineers are also using Argon agents for migrations from C and C++ to Rust. Google cited work ranging from smaller core libraries to more than 800,000 lines of code in the Zircon kernel used by the Fuchsia operating system.
Google said the resulting code still undergoes automated testing and human review before deployment.
Google compares Argon with OpenAI and Anthropic
Reuters reported that Google’s own evaluations place Argon ahead of OpenAI’s Astra and Anthropic’s Opus on some cybersecurity and other benchmarks, while it trails competing models on some coding tests.
Because Argon remains under restricted access, most external developers cannot yet independently test those performance claims. Initial third-party experience is coming mainly from Fairwind partners and other approved testers.
Google was also among several AI companies that signed the voluntary US safety agreement on September 29. The agreement includes commitments around pre-release evaluations and safeguards against unintended hacking or unauthorised access.
Argon will eventually be made available to enterprise and consumer customers, although Google has not committed to a public release date.
Following the initial cybersecurity testing period, the company plans to extend availability to paid API users and Google AI Ultra subscribers before a broader rollout.
Want to learn more about Cloud Computing from industry leaders? Check out Cyber Security & Cloud Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Tech Wire Asia is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
