Technology and Science30 September, 20263 min read
Google Halts Gemini 4 Argon Rollout Over Safety Concerns
This article is an automated aggregation of coverage from news outlets. Our editorials offer in-depth analysis curated by the NewsCord team.
Gemini 4 Argon rollout temporarily halted due to safety concerns Access limited to selected cybersecurity experts through Google's Fairwind program
25 outlets told this the same way.
Gemini 4 Argon debuts
Google announced Gemini 4 Argon as an AI model designed to handle complex workflows with a prominent role in cyber defense and autonomous detection of security vulnerabilities.
Google CEO Sundar Pichai signed, alongside U.S. President Donald Trump and other sector leaders, the White House Agreement on Superintelligence, a voluntary commitment driven from the White House to implement audits and internal controls in critical areas such as cybersecurity and biological threats.
Google said Argon topped the CWE-bench v1 benchmark, tied with another model, with a score of 68%, and Google said the system can autonomously locate, validate, and repair vulnerabilities in more than twenty programming languages.
Google said it would begin rolling out the model in a staged manner through its Fairwind program, granting participants access without the usual cybersecurity restrictions so they can leverage its defensive capabilities.
Google said Wiz, within its Scan for Good initiative, used Argon to discover a critical vulnerability exposing confidential personal information in software used in hospitals worldwide.
Access restricted, then questioned
Google temporarily restricted access to Gemini 4 Argon due to risks related to its potential misuse in cyberattacks, and Google said the model would be available only to a selected group of cybersecurity experts while the company gathered assessments of its capabilities and possible risks. Koray Kavukcuoglu, Google’s chief AI architect, wrote, “Safely launching frontier-level capabilities of this scale requires a gradual approach,” as Google described a phased rollout and voluntary early access for the U.S. government. Anonymous insiders with direct project access described difficulties of Gemini 4 Argon in some programming and interface design tasks, while Google rejected the assertion that the model performed worse at programming.
Google cited internal use and tests, including a case where Argon agents replaced 32,000 lines of SIMD code in libgav1 and produced a Rust decoder 2.7 times faster than the previous Rust version, with identical video output. Google announced an output limit of 1,000,000 tokens for Argon, compared with the previous 64,000.
Safety, pricing, and impact
Google said it would voluntarily grant the U.S. government early access to Gemini 4 Argon as part of a gradual rollout, and Google said it would use evaluators’ feedback to analyze the model’s behavior before making it available to the general public. Google stated that Argon includes restrictions against cyberattacks, and Google said the model is designed to reject requests that could facilitate the execution of cyberattacks or the creation of chemical, biological, or nuclear weapons. Argon would be made available “without cyber guardrails” to trusted defenders and internal Google teams so they can “leverage its full frontier-level cybersecurity defense capabilities.”
Argon’s output token limit is 1M tokens, up from 64K, and Argon would launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. Tulsee Doshi said starting the rollout with trusted cybersecurity partners gives “more confidence” and enables Google to put a model trained and strong in cyber defense “in the hands of defenders as soon as possible.”