Google has begun giving approved cyber defenders access to Gemini 4 Argon, a model the company says can find, validate and patch software vulnerabilities. Access starts with vetted Fairwind Program partners; Google says a wider release is planned, but it has not announced a date.
Fairwind is Google’s restricted early-access program for organizations protecting critical systems. Partners may use Argon for authorized threat simulation, reverse engineering and malware analysis for defensive or academic research. Google requires phishing-resistant multifactor authentication and limits access to internal cybersecurity, incident-response or penetration-testing teams.
Google says Argon will be available to these trusted defenders without cyber-specific model guardrails. That applies to the controlled Fairwind rollout: partner organizations must restrict access and may not use the model for malicious work such as creating malware.
On the public CWE-bench v1 leaderboard, Argon scores 68% on pass@1, tied for first with Grok 4.7 and GPT-6 Astra. The benchmark reports one-run (pass@1) and four-run (pass@4) results after four rollouts per model on each of 120 held-out audit-and-patch tasks; a passing patch must block the exploit and preserve all existing tests. These scores describe a defined coding test, not production reliability.
Google has not set a date for wider access through its paid API or Google AI Ultra.
Sources
- Google: Gemini 4 Argon announcement
- Google DeepMind: Fairwind Program
- Collinear AI: CWE-bench leaderboard and methodology
Image: Google data center in The Dalles, Oregon, by Lambtron, CC BY-SA 4.0 (illustrative).