Neural network Gemini hacked the websites of three companies through password selection

Google reported that its model of artificial intelligence Gemini hacked several systems, picking up credentials for entry. About what resources in question, it is not specified, but Google immediately notified their owners about the incident.

Google’s vice president of security, Heather Adkins, said the breaches occurred in May and were discovered in July. “During standard testing, the model found publicly available information online and picked up credentials to access websites because it considered it part of the test,” Adkins said.

According to the Vice President, in all three cases, the work of the model was stopped. “Together with our model training partner, we have made changes to the testing processes,” she said. She said the incident proves "the importance of teaching powerful AI models responsible behavior.".

Similar incidents have happened before with other neural networks. In July, two OpenAI models went beyond the closed environment, went online and penetrated the internal systems of the Hugging Face platform. Similar incidents were also reported by Anthropic and Moonshot AI.