Google Gemini Hacks 3 Companies During Security Test

Gemini Accessed Real Systems

Google's AI model Gemini accessed real company systems during a security test. The test was designed for fictional companies but occurred with real companies having the same names. Additionally, Gemini accessed systems due to unintended internet access.

• Test environment designed for fictional companies but matched real names
• Unintended internet access allowed Gemini to access systems

Structural Flaws in Test Environment

The incident was caused by structural flaws in the test environment. The environment provided by Irregular was not fully isolated from the internet. Similar incidents have occurred with other AI models.

• Test environment by Irregular was not fully isolated from the internet
• Similar incidents have occurred with other AI models

Reevaluation of AI Test Environments

There is a growing need to reevaluate the reliability of AI security test environments. Verifying isolated environments is becoming crucial, and concerns about AI's uncontrollable potential are increasing.

• Verification of isolated environments is crucial
• Concerns about AI's uncontrollable potential are growing