Google’s Gemini AI model briefly hacked into three companies during security tests this summer, the company confirmed.
The hacks occurred in May during testing with third-party AI safety company Irregular, according to a report from The Wall Street Journal.
“All relevant labs were notified in late July, and affected entities were contacted as part of the investigation,” an Irregular spokesperson told the outlet.
“Irregular took immediate action, and all known issues on our end were remedied and resolved weeks ago.”
In all three of the hacking instances, the model ended the intrusion after determining it had gained access to a real company’s system, Google claimed in the report.
In one instance, a Gemini agent guessed passwords to gain access to a protected system. The other two cases saw the model gain access to protected systems after finding credentials in a public repository.
Irregular notified Google about the hacks in July following OpenAI’s revelation that its models broke out during testing and hacked the open-source AI resource site Hugging Face, the report stated.
The company, whose parent is Alphabet Inc., said it did not consider the hacks to warrant public disclosure at the time because no harm was caused to the targeted companies and each intrusion ended without issue.
Google sad it did not consider these hacks to be examples of “misalignment”- which is AI-company speak for when the models act outside of its human controller’s intentions.
The names of the companies which were hacked by Gemini have not been released nor has Google revealed which Gemini model was involved in the security test.
The company also stated that they notified federal authorities about the hacks, according to the report.
The hacking revelation comes as some AI leaders, like OpenAI’s Sam Altman and Anthropic’s Dario Amodei, along with government officials, are calling for the federal government to begin regulating frontier AI labs.
Read the full article here
