Google Locked Its AI Vulnerability Hunter Behind 650 Seats. Build Your Own for Under a Cent per File

Google unveiled Gemini 3.8 Flash Cyber, an AI model capable of discovering critical vulnerabilities swiftly. This innovation is part of Google's Cloud Vulnerability Research team's efforts, which typically require more time. The Chrome Security team found that Flash Cyber produced 2.6 times more accurate patches for real Chrome vulnerabilities compared to the best commercial models.

However, there's no public access to the model - it's limited to a select group of approximately 650 organizations, including government entities, critical infrastructure operators, and select partners. Google's Chrome Security team utilized this model, achieving a 2.6x improvement in detection. The company's public model, priced at $0.75 per million input tokens, offers access to all.

The author, a backend engineer, built a two-layer audit pipeline using open-source scanners for known patterns and the public Flash model for logic flaws. On a test app with deliberate vulnerabilities, the combination of scanners and the model provided comprehensive coverage, catching issues that neither layer would have alone. Google's move to offer tiered access to such security-capable AI models alongside locked versions for external use reflects a cautious approach, balancing the potential for misuse with the need for robust security defenses.