The race to dominate artificial intelligence applied to cybersecurity has just intensified. Google has unveiled Gemini 3.5 Flash Cyber, a specialized variant of its AI model designed to automatically detect and resolve software vulnerabilities. With this move, the California-based company joins the direct competition with Anthropic's Claude Mythos and OpenAI's GPT-5.5 Cyber, two rivals that already dominated this strategic niche.

Faster and cheaper AI, but equally capable

Google DeepMind's strategy is particularly shrewd. By developing Gemini 3.5 Flash Cyber from its lighter base model, the company has achieved what seemed difficult: maintaining competitive performance without skyrocketing execution costs. This is crucial in a market where deploying these tools at scale can be prohibitively expensive.

According to hipertextual.com, Google argues that a more economical model opens previously unthinkable possibilities: CodeMender, the AI agent Google announced in 2025, could invoke Gemini 3.5 Flash Cyber multiple times to analyze the same large codebase. The result would be deeper analysis and frequent scans without costs becoming astronomical or processing bottlenecks appearing.

Numbers that speak for themselves

In performance tests, the numbers are convincing. On CyberGym, a benchmark specifically designed to measure how AI agents resolve real vulnerabilities, Gemini 3.5 Flash Cyber achieved an 83.2% success rate. To put it in context:

• Claude Mythos preview: 83.1%
• Claude Mythos 5: 83.8%
• GPT-5.6 Sol: 83.6%
• GPT-5.5 Cyber: 85.6%

Google is barely one percentage point below its closest competitors, demonstrating that it has not sacrificed capability for efficiency. It is a rare combination in technology: being as powerful as rivals, but running faster and costing less.

Already in use within Chrome, YouTube and Android

Google does not wait for end users to test its tool. The company already integrates Gemini 3.5 Flash Cyber in CodeMender to identify and fix vulnerabilities in its own products: Chrome, YouTube and Android. This allows it to validate the technology in real-world environments while improving the security of services used by billions of people.

Restricted access for now, expansion promised

For the moment, regular users will have to wait. Google will grant initial exclusive access to governments and strategic partners through CodeMender. The company promises to expand access over time, although it has not yet revealed specific timelines or exact conditions under which it will reach end users.

Meanwhile, Google has also announced Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, expanding its model catalog. Additionally, the "pretraining" phase of Gemini 4 is already underway, suggesting the AI race in cybersecurity is just beginning.

The question this move leaves is clear: can the balance between performance, speed and cost that Google offers reconfigure the cybersecurity market?