Cybersecurity AI "Too Risky to Release"—Are Chinese Open-Source Models Catching Up?

Google Gemini 4 Argon Offered to Select Partners First

GLM-5.3 Approaches Mythos in Some Attack Evaluations

AI Downloaded by Attackers Too... Access Control Alone Faces Limits

While American artificial intelligence (AI) companies are restricting access to high-performance AI models, Chinese open-source models are demonstrating similar capabilities in cyberattack evaluations, raising questions about the effectiveness of access controls. As the number of models available to potential attackers increases, it is expected that it will become more difficult to prevent the spread of risks through individual companies' access controls alone.

According to the AI industry on October 2, Google is providing its latest frontier AI, Gemini 4 Argon, first to trusted security partners through its cyber defense collaboration program "Fairwind." The company has not disclosed a specific timeline for when the general public will be able to use the model.

Argon is capable of handling complex tasks in software development, specialized work in legal and financial fields, and cyber defense. Google plans to expand availability after addressing safety based on feedback from initial partners. Rather than releasing high-performance models directly to the public, the company is selectively disclosing them and expanding access step by step.

Anthropic and OpenAI are also making their advanced AI models available only to select parties through programs like Glasswing and Daybreak. The idea is to give defenders time to find and fix vulnerabilities in major software before attackers can acquire similar capabilities.

However, these levels of cybersecurity capability are also being matched by open-source models. Anthropic recently reported that the open-weight model GLM-5.3 from China's Zhipu was able to successfully exploit vulnerabilities in Google Chrome 50 out of 410 attempts in the "ExploitBench" test. This is comparable to the 56 successful attempts achieved by Mythos Preview. The U.S. National Institute of Standards and Technology (NIST) stated, "Among all open-weight models released so far, this one demonstrates the most advanced cybersecurity capabilities," adding that "it is about four months behind the cutting-edge models produced in the United States."

The rapid narrowing of the cybersecurity gap by open-source models is making it more difficult for developers to control misuse simply by restricting access. Even if a particular company provides its models only to defenders, attackers can still utilize separate open-source models they acquire elsewhere. Anthropic commented, "Attackers may not be able to easily gain access to U.S. models, but GLM-5.3 is available for anyone to download," adding, "The lax safety features in GLM-5.3 are believed to significantly enhance the cyberattack capabilities of malicious actors."

As the offensive capabilities of open-source models increase, so too does the need to combine access restrictions with expanded defensive capabilities. Choi Byungho, Professor at Korea University's Human-Inspired AI Research Institute, said, "As Chinese models become more capable in cyberspace, attackers have increasing opportunities to access powerful models," and added, "It is necessary to strengthen the ability to detect abnormal behavior in real time using lightweight models, and to establish clear decision-making procedures and responsibilities for how far AI should respond once an attack is detected."

This content was produced with the assistance of AI translation services.