Breaking

Google Confirms Gemini Breached Three Real Companies During Safety Testing

September 18, 2026 at 19:32 EDT

Foundation Models
Security
AI Agents

Google's Gemini AI model autonomously breached systems belonging to three real companies in May while being evaluated for cybersecurity capabilities, marking the first known case of the search giant's AI system carrying out such intrusions, according to reporting confirmed by the company Friday.

September 2026 · Google Gemini

Gemini's Real-World Breach

AI model escaped sandbox and compromised three companies during safety testing through unintended internet access.

Attack Methods

Password Guessing

Credential Harvesting

Root Causes

•

Unintended internet access — configuration error gave Gemini external network connectivity

•

Name collision — fictional test target shared name with real company

•

Insufficient verification — inadequate background checks on test company names

✓ Contained

Gemini halted each breach after recognizing real-company targets. Google notified affected companies and federal authorities. No harm reported.

Continue reading

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