Artificial Intelligence

Google Will Not Sell You Its Best Bug Hunter

Gemini 3.5 Flash Cyber found more real vulnerabilities than Claude Opus 4.6, but Google is releasing it only to governments. Here is what that asymmetry means for you.

Artificial IntelligenceBy Servola Tech Desk2026-07-233 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Google Will Not Sell You Its Best Bug Hunter

Key takeaways

Google's new Gemini 3.5 Flash Cyber found 55 confirmed vulnerabilities in the V8 engine against 36 for Claude Opus 4.6, then Google restricted it to governments and trusted partners - so the best result is the one you cannot procure.
A small, cheap model beating a frontier model at real bug-finding is the signal: for narrow security work, specialization now outperforms raw scale, and it does so with 17 percent fewer output tokens.
The capability, not the product, is what changed. Automated vulnerability discovery is now demonstrably cheap, so plan on adversaries reaching a comparable tool before your defenders can license this one.
The edge you can actually buy this quarter is unglamorous: shorter patch latency, memory-safe languages and a real software bill of materials - not a defensive model held behind a government pilot.

A small model out-hunted a frontier one

On July 21 Google DeepMind released Gemini 3.5 Flash Cyber, a security-tuned version of its lightweight Flash model built to find, confirm and patch software flaws. Tested against the V8 JavaScript engine that runs inside Chrome, it surfaced 55 unique confirmed vulnerabilities, against 47 for the mainline Flash model and 36 for Claude Opus 4.6. Ten of them were issues neither of the other models caught.

The headline everyone will write is that a defensive tool got better. The number an owner should sit with is that a small, cheap model beat a large frontier one at finding real bugs in one of the most scrutinized codebases in the world.

The economics just inverted

For two years the working assumption was that the biggest, most expensive model wins. Flash Cyber breaks that for narrow tasks. It runs on a cost-efficient base, uses roughly 17 percent fewer output tokens than the model it was tuned from, and still leads on the CyberGym benchmark and on Chrome's production commit scanning. Specialization, not size, carried it.

That is a procurement signal before it is a security one. You do not need frontier-tier general models to scan your own code; a smaller model tuned for the job can be cheaper and better, which reframes how you budget for AI tooling across the board.

The catch is who gets to hold it

Google is not selling Flash Cyber. It is being released through a limited pilot to governments and trusted partners, integrated into Google's CodeMender agent, with a stated aim of giving frontline defenders a head start while limiting misuse of a dual-use capability. That reasoning is honest. It also means the strongest version of this tool is, for now, something you cannot license.

Google does offer CodeMender's foundational capabilities through its enterprise agent platform using standard models. Useful, but it is not the model that beat Opus, and you should plan around the gap rather than the promise.

Plan for the capability, not the product

Access controls slow diffusion; they do not stop it. The important fact is no longer which company holds the best scanner - it is that automated vulnerability discovery at this level has been demonstrated cheaply, and cheap capabilities get replicated in the open. Assume an attacker can reach a comparable tool, and build your risk model on that assumption rather than on Google's release policy.

For a European team, that lines up with where regulation is already pushing. The Cyber Resilience Act and memory-safety guidance from agencies like Germany's BSI and the UK's NCSC all point at the same durable defenses, which no model gate can take away from you.

What actually reduces your exposure

The unglamorous controls are the ones you own. Shorten the time between a patch shipping and you deploying it, because automated discovery shortens the attacker's side of that same race. Move new code toward memory-safe languages where the whole class of bugs Flash Cyber hunts simply does not compile. Keep a real software bill of materials so that when a flaw lands in a dependency, you know within hours whether it is in your stack.

None of that waits on a pilot programme. It is available today, it compounds, and unlike a withheld model, no one can revoke your access to it.

Frequently asked questions

Can my company use Gemini 3.5 Flash Cyber?

Not directly for now. Google is releasing the model through a limited pilot to governments and trusted partners rather than as a general product. You can access CodeMender's foundational capabilities through Google's enterprise agent platform using standard models, but that is not the security-tuned model that outperformed Claude Opus 4.6.

Does a tool like this make attacks against me easier?

Indirectly, yes. The result proves that cheap, automated vulnerability discovery works, and demonstrated capabilities tend to be replicated openly. The prudent assumption is that attackers will reach comparable tooling, which is why shortening your patch latency matters more now, not less.

If I cannot buy the model, what should I invest in instead?

The defenses you control outright: faster patch deployment, moving new code to memory-safe languages, and maintaining an accurate software bill of materials so you can trace a vulnerable dependency in hours. These compound over time and do not depend on any vendor's release policy.

The interesting news is not that Google built a better bug hunter; it is that the capability is now cheap enough to assume everyone will have it, which makes the defenses you own outright the only ones worth betting on.

Gemini 3.5 Flash CyberGoogle DeepMindCodeMendervulnerability discoveryClaude Opus 4.6dual-use AImemory safety

More Than 15 Partners, and No Published Rules

Google DeepMind says it advanced more than 15 biosecurity partnerships in 12 months and named none of them. There are no published criteria, no application process and no appeal path. That is a procurement fact, not an ethics complaint.

6 min read

Decisions

Google Scrapped a Near-Finished AI Model

Google DeepMind scrapped a nearly finished Gemini base model and moved the launch to 17 July. The sunk-cost discipline behind the call, and the trap it sets for buyers.

3 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.