Google says its Gemini AI model hackit three ither companies

In a first for Google, the company confirmed that its AI model, Gemini, breached the security of three ither companies back in May. The hacks cam aboot durin a cybersecurity evaluation by the AI-security firm Irregular.

Irregular, an Israel-based startup that sichts the security of advanced AI systems, was likewise at the centre o some o the recent OpenAI and Anthropic hacks o third-party entities, includin OpenAI’s breach o the AI software company, Hugging Face.

The circumstances that alloued the models to hack ither companies in some o these cases are verra similar: Irregular was testin the models in a closed environment wi fake companies. The testin environment wasna supposed to be internet-enabled, but access was made available unkent, accordin to the Wall Street Journal. Once connectit to the internet, the models unexpected-like hackit intae real firms.

Irregular telt Google o the hacks at the end o Julie efter discoverin that OpenAI had hackit intae Hugging Face. Google confirmt to the Guardian that the hacks did happen, but that the company didna feel it needit public disclosure acause the models didna skaith the companies. The Wall Street Journal reportit on the breaches and revealed for the first time that they had indeed occurred.

“In a standard evaluation, the model fand public information online and guessit credentials to access websites it thocht were pairt o the test,” Heather Adkins, vice-president o security engineering at Google, said in a statement. “In all three o these instances, the model stoppit.”

In one o the security breaches, Irregular was testin Gemini’s cybersecurity capabilities by promptin the AI model to get information from a fake company’s software. The fake company had the same name as a real company. When the model unintentionally gaint access to the internet, it richtly guessit the password o and breachit a real company’s service, Irregular telt the WSJ. Google said that aince it fand oot it had hackit a real company, and no the simulated one, it stoppit.

In twa ither tests, the model searchit the web for and fand public repositories containin credentials to twa ither companies. The model used thae credentials to access real companies. When it fand oot they were real companies, it stoppit, accordin to Google.

Anthropic and OpenAI chose to voluntarily disclose the hacks but Google didna. Houever, the company said it made siccar that the three companies that were hackit were made aware.

“These events highlight the importance o trainin pouerfu AI models to act responsibly,” Adkins, the Google spokesperson, said.

Anthropic and OpenAI’s disclosures promptit the independent senator Bernie Sanders to demand the companies pause development o their technology, sayin it signallit that the company was nae langer able to control their models.

OpenAI pausit development o their models for twa weeks, while Anthropic CEO Dario Amodei has cried for a collective slowdown o AI development to ensure that its maist advanced models are bein biggit wi enough safeguards.