This is a narrower release than a typical general-purpose model launch: Gemini 3.5 Flash Cyber is purpose-built for finding and helping patch security vulnerabilities in code, and during testing it reportedly surfaced fifty-five confirmed bugs in the V8 JavaScript engine, including around ten that general-purpose coding models missed entirely. That gap matters, because it suggests vulnerability discovery genuinely benefits from specialization rather than simply being a smaller slice of what a strong general coding model already does well. A model trained specifically to reason about memory safety, type confusion, exploitation primitives, and the particular ways security bugs tend to hide in large, complex codebases like a JavaScript engine appears to catch categories of issues that a broadly capable coding assistant, optimized for feature implementation and general code quality, simply isn't tuned to look for. The significant caveat here is access: this model isn't broadly available. It's currently limited to governments and select partners through DeepMind's CodeMender program rather than being offered as a general-purpose API anyone can call. For most developers, that means the direct, practical takeaway isn't "go use this tool," since you currently can't, but rather a signal about where specialized security-focused models are heading and roughly how much they can outperform general-purpose alternatives on this specific kind of task. It's also a reasonable indicator that as these specialized models mature and eventually become more broadly available, teams doing security-critical work on large legacy codebases may want to budget for a dedicated vulnerability-scanning pass with a specialized model rather than assuming their general-purpose coding assistant's security awareness is sufficient on its own.