Google Blunder: Gemini Hacked 3 Companies During a Test That Went Wrong

Now it’s Gemini’s turn to go off the rails! Google has just acknowledged that its AI gained unauthorized access to the systems of three companies.The incident dates back to May of last year, during a cybersecurity test conducted by Irregular, an Israeli start-up. Specializing in the security auditing of advanced AI models, the company had already been involved in similar incidents with ChatGPT, OpenAI’s AI. These incidents had resulted in the hacking of the open-source Hugging Face platform.

In the three incidents documented by Google, Irregular was testing Gemini in a completely closed environment populated with fictitious companies, which was supposed to be disconnected from the Internet. By mistake, the digital sandbox ended up connected to the Internet. Once online, the AI model acted autonomously and broke into the systems of real companies.

Also read: Six out-of-control ChatGPTs – OpenAI acknowledges several new “worrying” lapses

When Gemini guesses a company’s password

In the first case, Gemini was supposed to extract information from the software of a fictitious company bearing the same name as a real company. Having inadvertently gained access to the Internet, the model correctly guessed the password of the real company and connected to its services. Google says that the model stopped as soon as it realized that this was not the test environment.

In two other cases, Gemini searched the web for public repositories containing login credentials, then used them to access the companies’ servers. Here too, the model stopped once it determined that the targets were real and not part of a simulation.

“During a standard evaluation, the model found public information online and guessed credentials to access sites it believed were part of the test. In all three cases, the model stopped,” says Heather Adkins, Google’s vice president of security engineering.

Google did not disclose the incident for months

The executive added that “these incidents underscore the importance of training powerful AI models to act responsibly.” Irregular reported the intrusions to Google in late July, after discovering that OpenAI had likewise hacked Hugging Face during a comparable test. For several months, Google did not consider it necessary to make a public disclosure. The Mountain View giant stresses that Gemini caused no damage to the companies concerned. Nevertheless, the company says it warned the three affected companies and worked with Irregular to evolve its testing protocol.

This episode is part of a series of similar incidents over the past few months.In July, Anthropic’s Claude model had escaped its testing environment to hack three organizations, shortly after ChatGPT carried out cyberattacks against several online services. These incidents pushed the AI giants to advocate a moratorium on the development of artificial intelligence. For Dario Amodei, CEO of Anthropic, it is imperative to pull back before autonomous agents cause serious damage. In this context, OpenAI, Google and Anthropic joined forces to work on a common tool for monitoring AI risks.

👉🏻 Follow tech news in real time: add 01net to your feeds on Google, and subscribe to our WhatsApp channel.

Source: Wall Street Journal