Google gives vetted defenders a Gemini 4 Argon build without cyber guardrails

Google ties the unguarded build to a staged rollout that starts with trusted defenders in its Fairwind Program ^[1]. It says a phased approach is required to release capabilities at this level safely ^[7]. Before the broad release, Google says, it is strengthening safeguards against cyber and chemical, biological, radiological and nuclear misuse, and internal and external red teams have tested them ^[5]. Monitors watch Argon's reasoning and actions and stop execution when needed, according to Google ^[6]. The company is also taking part in the U.S. government's voluntary process for pre-release model access ^[8].

For organisations outside the vetted group, the launch does not change what an attacker can do today ^[3]. Help Net Security's account describes no attacker use of Argon and no copy of the unguarded build outside the Fairwind defenders and Google's internal teams ^[3]. On that record, the route to Argon for anyone else is the later public release, behind guardrails Google says it is still adjusting with feedback from early testers ^[7].

The long unattended runs come from a larger output budget. Google raised the output limit from 64,000 tokens to 1 million ^[11], about 15.6 times the old ceiling ^[1]. Google says the model can now think and write for hundreds of thousands of tokens in a single run ^[12]. A run that fills the full million-token window costs $10 in output at launch rates and $20 once they double ^[2]. Cached input costs $0.10 per million tokens during the introductory period ^[3].

Google's discovery numbers come from an internal test spanning complex codebases in 20 programming languages ^[16]. On fixing, the public evidence is one benchmark score and Google's own code migrations. Argon's CWE-bench v1 score is 32 points short of a perfect mark ^[4]. Argon agents replaced 32,000 lines of SIMD code in the libgav1 video decoder with Rust ^[13]. The new decoder runs 2.7 times faster than the earlier Rust port, with identical video output ^[13]. Agents also applied memory optimizations across Google's data centers that freed more than 300 TiB once rolled out ^[20].