Gemini's New Cyber Model Found a 13-Year-Old Chrome Bug

Google's Chrome security team used a new Gemini model to find a bug that had gone unflagged in Chromium for 13 years, the company said as it introduced Gemini 3.8 Flash Cyber on Sept. 2.

The model is Google's third Flash release in six weeks, following Gemini 3.7 Flash three weeks earlier and arriving a day after Google Pics went generally available for Workspace subscribers. Unlike the standard 3.8 Flash, which is priced at $0.75 per million input tokens and $3.75 per million output tokens and open to any developer, the Cyber variant is going only to "trusted defenders" through a new Fairwind Program, prioritizing government agencies, critical-infrastructure operators and maintainers of widely used software.

A bug dozens of engineers had missed

Chrome engineering director Doug Turner said the flaw had sat in Chromium and Chrome for 13 years, according to VentureBeat. He called it a "very subtle bug" that "dozens, if not hundreds, of engineers looked at but never flagged."

Turner described a "vulnerability apocalypse" driven by generative AI, saying reports through Chrome's own research program jumped in a "hockey stick increase." Google said 3.8 Flash Cyber produced 2.6 times as many correct patches for Chrome vulnerabilities as the larger commercial models it tested.

Benchmark scores, and one skeptic's framing

On CyberGym, an industry benchmark for autonomous vulnerability discovery, Flash Cyber scored 86.2%, up from 77.5% for the prior 3.5 Flash Cyber model. On CWE-Bench, a patching benchmark run by Collinear, it reached a 47.2% pass rate against 47.8% for a larger frontier model, at a lower cost per rollout, according to Google and VentureBeat.

Wiz, the cloud-security firm Google bought for $32 billion, ran its own penetration-testing benchmark against the model and reported a 7.5 to 9.7 percentage-point recall improvement over rival frontier models, at roughly one-fifth to two-fifths of the cost. Google's Cloud Vulnerability Research team said it found a critical foundational flaw in under two hours with the model, work that typically takes months.

Evaluation	Gemini 3.8 Flash Cyber	Comparison
CyberGym	86.2%	77.5% (3.5 Flash Cyber)
CWE-Bench	47.2%	47.8% (a larger frontier model)
Chrome patch accuracy	2.6x	vs. larger commercial models

Gemini Security Lead Raluca Ada Popa said the model was built to give defenders an edge, since "attackers need only find one significant flaw over millions of lines of code," while defenders must remove every one to stay safe. Google said 3.8 Flash Cyber ships with a more permissive set of cybersecurity mitigations than the standard model, which is why access is restricted rather than public.

The gated rollout is a different approach than OpenAI's for GPT-6 Astra, a model OpenAI rated at the top tier of its own cybersecurity risk scale yet shipped broadly, relying on monitoring safeguards rather than a closed defender program. Google has not said how many of the organizations that apply through Fairwind will actually receive Flash Cyber access rather than the standard model.