Generative artificial intelligence companies are increasingly giving their most capable models to security experts first and releasing them to the public only later. The shift reflects a concern that AI has advanced beyond merely finding software vulnerabilities to carrying out actual attacks, and that defenders need time to prepare before wider release.
Google released Gemini 4 Argon, its top-tier AI model, to the public on the 13th, according to information technology industry sources on the 15th. Argon is designed to solve complex, multi-step tasks across fields including coding, finance, law and cybersecurity.
Google first made Argon available on Sept. 30 to select experts participating in its cybersecurity partnership program, Fairwind, and to the U.S. government. The reason was that the model's ability to detect vulnerabilities could be turned to cyberattacks. The company strengthened the public version's resistance to prompt injection, an attack that manipulates AI through malicious instructions.
Other companies have taken the same approach. Anthropic in April released Claude Mythos Preview, an early version of its most advanced model, Mythos, to a limited group of companies. In June it released Fable 5, a Mythos-class model refined for safe general use, and gave Mythos 5, a security-focused model, to its cybersecurity partnership Project Glasswing. When it released Claude Mythos Preview in April, the company said it had judged the technology could be abused by hackers and others, and was therefore providing it first to major technology and security firms for defensive purposes.
OpenAI likewise gave GPT-6 Astra in September to select organizations taking part in its cybersecurity program, Daybreak, before expanding access to paying users over several days. Astra was also the first model to receive the highest risk rating, "critical," in OpenAI's own evaluation. At that level, a model given certain access privileges can identify security vulnerabilities and design and execute an attack without human intervention.
Behind these staggered releases is the rapid improvement in AI's offensive capabilities. Where AI once served only to assist with vulnerability analysis or code writing, the latest high-performance models can discover flaws themselves and mount attacks.
Companies are therefore choosing to put powerful models in defenders' hands first, buying time to analyze emerging attack techniques and devise countermeasures. "AI has developed to the point where it can automate everything from finding a vulnerability to executing an attack, so companies want to secure time to build defenses before public release," said Hwang Seok-jin, a professor of cyber forensics at Dongguk University's International Graduate School of Information Security.
He cautioned, however, that staged release does not guarantee safety. "Releasing a model in stages does not by itself ensure safety," Hwang said. "It has to be accompanied by user verification, access controls and real-time monitoring for abuse."