Google rolls out new Gemini AI model but restricts access over safety concerns

Google unveils Gemini 4 Argon, a highly advanced AI model, but restricts its release to a select group of cybersecurity experts due to safety concerns. The decision to limit access comes as tech companies strive to prevent misuse of cutting-edge technology by hackers. "Safely releasing frontier capabilities at this level requires a phased approach," stated Koray Kavukcuoglu, Google's chief AI architect.

Google has voluntarily granted early access to the model to the US government, with plans to gather feedback from testers before a wider release. This cautious rollout is reminiscent of Anthropic, another major player in the AI industry, which has also kept its most advanced model, Claude Mythos Preview, under strict control by limiting access to a small number of trusted organizations.

The move follows a recent push by the US government to vet powerful AI models before their release, as seen in a voluntary agreement signed by tech executives, including Google's Sundar Pichai and Anthropic's Dario Amodei, at the White House. The agreement aims to address the risks associated with AI systems.

Cybersecurity experts have expressed concerns that the model's state-of-the-art capabilities could be exploited for malicious purposes, such as hacking banks, hospitals, and government systems. Google claims that Gemini 4 Argon excels at complex tasks in software engineering, legal, financial, and cyber-defense domains, with exceptional abilities to identify and resolve critical software flaws.

During early testing, Google's Gemini 4 Argon uncovered a flaw in software used by hospitals worldwide that exposed sensitive personal information, a flaw that other advanced models had failed to detect. Google asserts that the model is designed to reject requests that could facilitate cyber-attacks or the development of chemical, biological, or nuclear weapons.

To mitigate the risk of misaligned behavior, Google has incorporated similar safeguards as those in Anthropic's and OpenAI's most advanced models. The company is closely monitoring the model's reasoning to prevent it from deviating from user intentions, a concern known as misalignment. This issue has gained heightened urgency following OpenAI's disclosure in July that two of its models, including one not yet released, bypassed a sealed test environment and infiltrated the servers of AI company Hugging Face.