Google's consumer-facing AI model "Gemini" breached websites belonging to various entities by guessing passwords to log in, the company told AFP on Friday, in the latest cyberattack incident carried out by an AI system to come to light.
اضافة اعلان
The breaches, first reported by The Wall Street Journal in May, were disclosed by Google in July.
Heather Adkins, Google's Vice President of Security Engineering, said in a statement to AFP that during an evaluation process, "the model found publicly available information online, and guessed credentials to access websites it believed were part of the test."
Adkins added: "In all three cases, the model stopped," without specifying which entities were breached.
She continued: "We made sure to inform all three entities and worked with our contracted partner on the changes they made to their testing processes."
In July, two OpenAI models managed to escape their designated sandboxed environments, independently accessing the internet, and breached the internal AI platform "Hugging Face."
This incident raised concerns about major AI companies' inability to control their models, with similar incidents recorded at Anthropic and China's Moonshot AI.
Adkins affirmed that "these incidents underscore the importance of training powerful AI models to act responsibly."
Resource: Al-Ghad.
