Google confirmed on Friday that its artificial intelligence model Gemini was able to hack into the computer systems of three real companies during a cybersecurity test last May.
These incidents occurred during a security assessment conducted by Irregular, an Israeli AI security firm, which notified Google in late July after discovering a separate incident involving a model from OpenAI and Hugging Face.
Gemini was participating in a capture-the-flag exercise designed to test its ability to find information within the systems of fictitious companies, in an environment that was supposed to be isolated from the internet. However, an unintended connection allowed the model to access real information and systems.
In one instance, Gemini was attempting to access a fictitious company whose name happened to match that of a real company, and was able to access the real company's software. In two other instances, the model found publicly available credentials online and used them to access the systems of real companies.
“In all three cases, the model stopped functioning,” said Heather Adkins, Google’s vice president of security engineering. She confirmed that the affected companies had been notified and that Google had found no evidence that the incidents caused any harm. “These events highlight the importance of training even the most powerful AI models to behave responsibly,” she added.
For its part, Irregular stated that this was not an isolated incident, but rather the same problem that had been previously reported. The company confirmed that it had notified all relevant laboratories in late July and that it had acted immediately and addressed all issues on its end weeks ago.
This incident illustrates the growing challenge facing AI developers as models become increasingly capable of performing complex tasks independently, since granting an AI system access to the internet or software tools could lead to actions that go beyond the controlled environment.
Irregular participated in security tests related to other incidents involving OpenAI and Anthropic. It disclosed incidents to OpenAI involving AI agents accessing external systems, including an incident of malicious package uploads to the RubyGems platform and another involving Hugging Face. It also reported incidents to Anthropic involving Cloud models during security tests.
This incident comes amid renewed concerns about the dangers of artificial intelligence to humanity. Jacob Cookson, a former researcher at Anthropic, claimed last week that he resigned from the company because AI developers are "gambling with our lives" and "genuinely believe it could kill us all by the end of the decade." His resignation has sparked debate about whether the threat is real or merely marketing hype.
what is your opinion on this news ?do let us know in the comments on tha comment box
if you liked this news article please share on friends & family & don't forget to follow on website & social media