Google rolls out Gemini 4 Argon to AI cyber defenders

On 1 October 2026 Google released Gemini 4 Argon, a security-specialised variant of its Gemini 4 model, to a limited group of trusted cyber defenders enrolled in its Fairwind Program. According to Google, Argon is designed to autonomously discover software vulnerabilities, validate that they are genuinely exploitable rather than theoretical, and generate patches — closing a loop that traditionally requires separate human teams for finding, triaging and fixing bugs. The company said Argon scored 68 percent on CWE-bench v1, a public benchmark that measures how reliably AI systems can identify and remediate issues drawn from the industry's Common Weakness Enumeration catalogue, a notable jump above previous general-purpose models. Google positioned the launch as part of a broader push to put advanced generative models squarely in the hands of defenders before they are weaponised by attackers, pairing the release with new disclosure guidelines and sandboxed evaluation infrastructure. The announcement arrived on the same day that OpenAI said it had warned more than 100 organisations about unauthorised activity tied to its agents, underscoring the escalating role of AI on both sides of cybersecurity.