Breaking

Google DeepMind launches Gemini 3.8 Flash Cyber for gated vulnerability detection and patching

September 2, 2026 at 16:58 EDT

Foundation Models
Security
AI Agents

Google DeepMind on Wednesday introduced Gemini 3.8 Flash Cyber, a cybersecurity-tuned model it says can autonomously find software vulnerabilities and generate validated patches, with access limited to vetted defenders through a new Fairwind Program.

September 2, 2026 · Google DeepMind

A defender-only model that finds and patches

Gemini 3.8 Flash Cyber is Google's most capable cybersecurity model yet: autonomous vulnerability discovery plus validated patches, gated through the Fairwind Program.

86.2%

CyberGym Pass@1

71.0%

20-language discovery

2.6×

Chrome valid patches

47.2%

CWE-Bench Pass@1

Magnitude

Chrome Security: 2.6× more correct patches

1×

Best larger commercial models

2.6×

Gemini 3.8 Flash Cyber

Chrome Security recorded 2.6 times more correct patches than the best larger commercial models tested.

Comparison · Pass@1

Real-world discovery across 20 languages

46.6%

3.5 Flash Cyber

58.9%

3.7 Flash

71.0%

3.8 Flash Cyber

Internal recall over 1,200 confirmed historical bugs. Column height is proportional to Pass@1.

How it is meant to work

CodeMender loop with Gemini 3.8 Flash Cyber

1 · Discover

Scan codebases for weaknesses

→

2 · Verify

Confirm the flaw is real

→

3 · Patch

Write a deployable fix in minutes

What testers like

Partners including Wiz, Palo Alto Networks, Snowflake, and Armadin report frontier-class hunting at Flash cost and latency, with less triage noise.

What remains contested

Access is restricted. Some engineers warn a bad auto-patch can cause an outage, and that attackers already have AI of their own.

Continue reading

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