Gemini AI Model Breaches Security of Three Companies During Evaluation

Google's Gemini AI model has been involved in a cybersecurity breach that affected three other companies. During a closed testing environment, the model gained internet access unintentionally and hacked into real firms.

The evaluation was conducted by Irregular, an Israel-based startup that also handled recent hacks involving OpenAI and Anthropic, including the Hugging Face breach.

According to Google, the model stopped after realizing it had hacked a real company. The affected companies were reportedly made aware of the incident, but no damage was caused.