Breaking

Google launches Gemini 3.8 Flash Cyber after Chrome tests show 2.6x more correct patches

September 2, 2026 at 12:16 EDT

Foundation Models
Security
AI Agents

Google on September 2, 2026 introduced Gemini 3.8 Flash Cyber, a cybersecurity-specialized model it says produced 2.6 times more correct patches on Chrome vulnerabilities than much larger commercial models, while pairing a general-purpose Gemini 3.8 Flash workhorse with restricted access for trusted defenders. The company is already using the cyber variant internally and is routing external access through a new limited program rather than a public API.

Sep 2, 2026 · Google DeepMind

2.6x more correct Chrome patches

Gemini 3.8 Flash Cyber, Google’s restricted cybersecurity model, produced 2.6 times more correct patches on Chrome vulnerabilities than much larger commercial models, with Flash-class speed and gated Fairwind access.

Correct Chrome patches vs larger commercial models

Larger
commercial

2.6x

3.8 Flash
Cyber

Chrome Security: more correct patches at Flash-class cost, not a larger general model.

86.2%

CyberGym pass@1

47.2%

CWE-Bench pass@1

70%+

Internal vuln discovery (20 languages)

2.3–5.2x

Lower cost vs frontier (Wiz)

1 Detect vuln

→

2 CodeMender verify

→

3 Patch in minutes

Defender view

Partners say Flash-level price and latency match larger models on pen-test and web-security tasks; Google Cloud research found a critical flaw in under two hours.

Limits

Not public: Fairwind-only, more permissive cyber mitigations, MFA and team-only use. CWE-Bench still trails the top frontier score (47.8%) by 0.6 points.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.