Google Launches Leaner, Cheaper AI Security Alternative

Google has unveiled a new AI security model, Gemini 3.5 Flash Cyber, specifically designed for rapid identification and remediation of security vulnerabilities. In a recent blog post, the company positioned Gemini 3.5 Flash Cyber as a “cost-efficient and highly capable alternative” to more extensive and costly AI systems, notably referencing Anthropic’s Mythos.

Constructed upon the foundation of Gemini 3.5 Flash, this new model will initially be accessible to governmental bodies and select trusted partners through CodeMender, Google’s specialized security coding agent. Google highlights that CodeMender can leverage 3.5 Flash Cyber repeatedly at speed and minimal expense, enabling the AI agents to analyze a broader range of code paths and uncover vulnerabilities more comprehensively.

In the competitive landscape, Anthropic’s Mythos 5 stands as a formidable AI security model, launched under its Project Glasswing initiative. This compute-intensive model carries a significant operational cost, reportedly double that of Claude Opus 4.8. Microsoft, having integrated Mythos for its security assessments, experienced its most substantial Patch Tuesday recently, attributing the success to AI-driven vulnerability detection. Amidst this, companies such as Google are striving to innovate, while China’s Z.ai has asserted that its own model possesses capabilities comparable to Mythos.

Google reported that 3.5 Flash Cyber demonstrated “competitive performance” against “significantly larger models” on the CyberGym AI cybersecurity benchmark, particularly when invoked up to five times. The model successfully identified 55 “unique confirmed issues” within the V8 JavaScript Engine, surpassing the 47 found by Gemini 3.5 Flash and 36 by Opus 4.6. Furthermore, Google noted that 3.5 Flash Cyber uncovered 10 issues previously undetected by any other model, highlighting its ability to continuously discover new code paths and vulnerabilities through multiple invocations.