Google's AI model Gemini hacked into three companies during cybersecurity test
This incident is the first known case of the model carrying out a security breach of this nature autonomously
Google’s AI model Gemini hacked into three companies’ computer systems during a cybersecurity test in May, the company has confirmed.
Listen to this article
The hacks were carried out during a cybersecurity evaluation, according to Google.
The incident is believed to be the first known example of the firm’s AI autonomously carrying out hacking of this nature.
During the security test, Gemini found publicly available information online and used it to guess credentials to access three websites.
Heather Adkins, Google’s vice president of security engineering, said the AI model believed the websites were part of the parameters of the test.
Read more: Are AI companies building hell and calling it success?
"We ensured the three entities were made aware, and we worked with our training partner on the changes they’ve now made to their testing processes," Adkins said.
"These events highlight the importance of training powerful AI models to act responsibly."
Irregular, the company which carried out the security test, has also been linked to OpenAI, Meta and Anthropic, including OpenAI’s tech platform Hugging Face.
Gemini was undergoing a test of its capability to locate information inside the systems of fictional companies, according to the Wall Street Journal.
The conditions under which Irregular was carrying out the tests were not meant to be internet-enabled, but the system was accidentally connected to the internet, the Wall Street Journal reported.
This allowed Gemini to subsequently hack into three other real firms.
An Irregular representative said other AI labs were affected by the same issue, and all relevant labs were notified in late July.
"All known issues on our end were remedied and resolved weeks ago," the spokesperson said.
The series of hacking incidents involving AI models has highlighted the need for safeguards for these platforms as they gain access to the internet and computer systems.
Anthropic boss Dario Amodei made headlines last week after his warnings concerning the rate of AI development.
Google’s vice president of security engineering added that in all three events, Gemini stopped its hacking.