Google reserves Gemini 4 Argon for cybersecurity partners

Google has unveiled Gemini 4 Argon, the most advanced AI model to date, with access initially limited to cybersecurity partners. This new flagship model in Google's Gemini series is designed to excel in software engineering, complex professional work, and cyber defence tasks, aiming to compete with OpenAI and Anthropic. The September 30 announcement marks a significant step forward for Google's AI capabilities, though it does not provide a timeline for broader availability, as the company plans to collect feedback and enhance safeguards before expanding access.

Cybersecurity experts are among the first to receive Argon through Google's Fairwind program. The company is also participating in the US government's voluntary process for model access prior to public release, emphasizing security evaluation throughout the rollout. Koray Kavukcuoglu, Google's chief AI architect and senior vice-president at Google DeepMind, highlighted that Argon is engineered to maintain reasoning through lengthy software development, legal, financial tasks, and cybersecurity challenges.

Google asserts that Argon outperforms OpenAI's Astra and Anthropic's Opus on crucial coding and cybersecurity tests, describing it as larger than its previous advanced Pro models and on par with its competitors. While Google's published benchmarks show Argon leading in several areas, the performance results vary across different tests, indicating that Argon excels in some professional AI applications but may fall short in others.

During testing, Argon achieved 77.9% on DeepSWE v1.1, a software engineering task, compared to 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra. In financial research and legal drafting, Argon scored 51.3% and 91.7%, respectively, indicating strong capabilities in these domains. Additionally, Argon's maximum output capacity has been increased to one million tokens, a substantial leap from the previous limit of 64,000, allowing for more extensive and complex reasoning in a single processing run.

Google emphasizes that cybersecurity is the immediate priority for Argon, with the model capable of autonomously identifying, validating, and repairing critical software vulnerabilities. Trusted cybersecurity professionals and internal Google teams will have unrestricted access to enable full utilization of its defensive features.

A cybersecurity company, Wiz, has already employed Argon through Scan for Good, where it detected a significant vulnerability in healthcare software used globally, exposing personal information. This vulnerability exemplifies Argon's potential to identify and mitigate serious security flaws.

Before the broader release, Google plans to reinforce protections against malicious cyber activities, including chemical, biological, radiological, and nuclear threats. The company is also addressing prompt injection vulnerabilities, where hostile instructions embedded in AI inputs attempt to manipulate its behavior. Additional controls monitor the model's reasoning and actions, capable of halting execution if it deviates from user intentions.

Internal and external specialists have rigorously tested these safeguards through simulated cyber attacks, with Google further isolating its testing environments to ensure security.