Google Restricts Gemini 3.5 Flash Cyber to Governments and Trusted Partners

Google is rolling out a new specialized security model, Gemini 3.5 Flash Cyber, built to find and patch software vulnerabilities faster and cheaper than the large, expensive AI systems that currently dominate the vulnerability-hunting field, The Verge reported on July 21. The model arrived alongside two general updates, Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, but it is the Cyber variant that Google is positioning as a direct answer to costlier rivals such as Anthropic’s Mythos. Rather than opening the model to the public, Google is restricting it to a limited pilot program for governments and trusted partners, citing its dual-use nature, according to The Hacker News.

A lightweight model built to run inside CodeMender

Gemini 3.5 Flash Cyber is not a standalone product. It runs inside CodeMender, the AI-powered vulnerability discovery and patching agent Google unveiled in October 2025. Google describes the new model as a “cost-efficient and highly capable alternative” to larger, more expensive cybersecurity-focused systems, built on top of the standard Gemini 3.5 Flash architecture rather than a heavier custom model. That lighter footprint is the point: CodeMender can call on 3.5 Flash Cyber “multiple times at high speed and low cost,” letting the agent scan far more code paths per vulnerability hunt than a slower, pricier model would allow on the same budget. For comparison, The Verge reports that Anthropic’s Mythos 5, a compute-heavy security model released as part of the company’s Project Glasswing initiative, costs twice as much to run as Claude Opus 4.8.

Outscoring bigger models on real codebases

Google’s own testing suggests the smaller model is not just cheaper, it also finds more bugs. On the CyberGym AI cybersecurity benchmark, Google says 3.5 Flash Cyber achieved “competitive performance” compared with “significantly larger models” when invoked up to five times per task. In a stress test against the V8 JavaScript engine, the model turned up 55 unique confirmed issues, ahead of the 47 found by standard Gemini 3.5 Flash and the 36 found by Claude Opus 4.6, including 10 issues that neither of those models caught at all. In a separate evaluation targeting complex, widely used codebases including Chrome and Safari, Google’s own announcement post says 3.5 Flash Cyber “significantly” surpassed mainline Gemini 3.5 Flash and 3.6 Flash, with a similar uplift from 3.5 Flash Cyber over 3.5 Flash on Chrome’s own production commit-scanning pipeline; Google notes that Claude models newer than Opus 4.6 declined those particular tasks under their own safety guardrails, so they were not part of that specific comparison. In one case, Google says, the model produced a 100% reliable remote code execution exploit that bypassed standard mitigations, including Address Space Layout Randomization (ASLR) and Write XOR Execute (W^X) protections, the same class of defenses security teams rely on to blunt memory-corruption bugs.

Why Google is not opening the model to everyone

That offensive capability is exactly why Google is not shipping 3.5 Flash Cyber broadly. “Given the dual-use nature of this technology, we have taken an intentional approach to how we deploy 3.5 Flash Cyber,” Raluca Ada Popa, DeepMind’s Gemini Security Lead, and Four Flynn, the company’s vice president of security and privacy, said in a blog post shared with The Hacker News ahead of publication. “As part of a limited-access pilot program, 3.5 Flash Cyber will be exclusively available to governments and trusted partners via CodeMender, expanding over time. This will give frontline defenders a head start in finding and fixing critical vulnerabilities before they can be exploited, while mitigating against broader misuse.” Because the model runs solely inside CodeMender rather than as an open-ended chat assistant, Google can set narrower guardrails: the system enables the model’s defensive functions while disabling other cyber-related activity, which also sidesteps a separate problem where an overly cautious model refuses to help a legitimate defender with forensic analysis. A DeepMind spokesperson told The Hacker News the company plans to extend the model’s capabilities to red-teaming and end-to-end enterprise defense over time, and Google is separately bringing CodeMender’s underlying capabilities to enterprise customers through its generally available Gemini models via the Gemini Enterprise Agent Platform.

A widening race to automate vulnerability hunting

The launch lands in the middle of an industry-wide push to put AI-assisted bug hunting into production. The Verge notes that Microsoft, which adopted Mythos for its own security checks, had its biggest Patch Tuesday to date this month after using AI to help find vulnerabilities, and that China’s Z.ai has claimed its own model can compete with Mythos’s capabilities. For now, Google is betting that a cheaper, narrowly deployed model can match the results of costlier rivals without the same price tag, while keeping the riskiest capability behind a pilot program limited to governments and vetted partners rather than the open market.