55 Bugs to Claude's 36 — Google's Gemini Cyber AI Works, and You Can't Have It

The tool works

On July 21st, 2026, Google DeepMind released a security model called Gemini 3.5 Flash Cyber. Pointed at Chrome’s V8 engine — one of the most scrutinized codebases on the planet — it found 55 confirmed vulnerabilities. Claude Opus 4.6 found 36. Ten of the flaws it surfaced, no other model caught at all.

That is not a demo number. V8 is audited constantly by some of the best security engineers alive, because a flaw there is a flaw in a couple billion browsers. Flash Cyber found real, exploitable holes anyway. Built on Google’s cheap, fast Flash model, it can scan millions of lines continuously without a frontier model’s bill. For defenders, who have always been outnumbered, that economics shift is the genuine promise.

Why you can’t have it

The same output, in a different pair of hands, is a zero-day factory. Give the model to a defender and it patches Chrome. Give it to an attacker and it writes a list of ways into Chrome that no patch exists for yet. Same model, same result, opposite consequence.

So Google gated it. Flash Cyber isn’t on the price list — it goes to governments and a short roster of trusted partners through a limited pilot. For almost every organization, the honest answer to “who is this for” is: not you, and you can’t get it. The capability is being rationed by a private company, on that company’s own definition of who counts as trusted.

The pattern that matters

I’ve spent 25 years in cybersecurity, and the defender-attacker symmetry is the oldest story in the field. Every capability shows up for both sides at once. What’s new is the response. One month before this, the U.S. Commerce Department ordered Anthropic to stop exporting its Mythos and Fable models over the same class of capability. Two of the largest labs in the world, one month apart, concluding that a cyber-capable model can’t just be sold to whoever wants it.

The benchmark was always going to climb. The line we just crossed is that the capability itself now decides who’s allowed to hold it — and right now the people drawing that line are the labs, not any public rule. The tool got locked in a vault. The question is who holds the key, and nobody voted for them.