Google limits Gemini 4 Argon to trusted defenders

Google is offering Gemini 4 Argon first to vetted defenders via Fairwind, with conflicting reports on guardrails, tokens, and a hospital flaw.

Google DeepMind announced Gemini 4 Argon, a frontier model for software engineering, enterprise work, and cybersecurity, initially limited to trusted defenders and internal teams through the Fairwind Program. Reports say it can find, validate, and patch vulnerabilities and that it beat Gemini 3.8 Flash Cyber on attack-surface discovery and proof-of-concept generation. Sources disagree on guardrails: The Hacker News says a guardrail-free version is planned, SecurityWeek says Fairwind access starts without cyber guardrails, Security Affairs says restrictions are relaxed for approved teams, and Google says it is still collecting guardrail feedback while monitoring chain-of-thought so execution can be halted. Token claims also conflict—outlets report a 1 million token output cap, up from 64,000, while Google's recap describes a 1-million-token context window—and introductory pricing is given as $2 per million input tokens and $10 per million output tokens, later $4 and $20. Cited scores include 68% on CWE-bench v1, with SecurityWeek naming a tie with GPT-6 Astra and Grok 4.7 and CSO Online also including Claude Opus 5.5, plus 77.9% on DeepSWE v1.1, 91.7% on LVBench, 51.3% on AutomationBench, 68.9% on the Vals Index, and a Gray Swan indirect prompt injection lead; Claude Opus 5.5 leads PostTrainBench at 49.3% to Argon's 45.3%. Reports describe a critical flaw in worldwide hospital software, but disagree on whether Argon or Wiz found it and whether personal-data exposure was established.