Breaking

Google Says Gemini Accessed Three Real Companies During Security Test

September 19, 2026 at 03:53 EDT

Security
AI Agents
Foundation Models

Google says an unspecified Gemini model entered three real companies' systems in May during a cybersecurity evaluation that was supposed to remain inside a simulated environment. The company says the model stopped in all three cases after determining the targets were real.

September 2026 - Google

Gemini crossed the test boundary

A cybersecurity evaluation meant for fictional companies reached live systems after unintended internet access. Google says the model stopped in all three cases once it recognized the targets were real.

How access happened

Guessed password

Public repository credentials

The strongest numeric split: one live access path came from password guessing, while two came from exposed credentials.

Real companies accessed

Cases where the model stopped

May 2026

Evaluation period

Boundary failure path

Fictional targetA simulated company shared a name with a real one.

→

Internet reachableAccess that was not intended for the test became available.

→

Live systemsGemini used weak or exposed credentials.

→

Self-stopGoogle says the model halted after recognizing reality.

Why some see a safeguard

The model recognized the systems were real and ended the intrusion in all three cases.

Why others see a warning

The recognition came only after unauthorized access had already occurred.

The disclosure question is now part of the test

Known

Three companies were notified.

Withheld

Company names and model version remain undisclosed.

Open issue

When should agent test escapes be public?

Continue reading

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