Google’s artificial intelligence model, Gemini, penetrated external computer systems by guessing credentials before stopping on its own, the company said to AFP on Friday, confirming a report from The Wall Street Journal.
In May, “during a standard test (of Gemini), the model found public information online and guessed credentials to access computer sites that it thought were part of the test,” explained Heather Adkins, one of Google’s security executives.
Three different organizations were affected, according to Google, which did not name them but said it had informed them of these breaches.
In one case, according to The Wall Street Journal which revealed the information, “the model tested different password combinations until it managed to access a protected system.”
“All three times, the model stopped,” reported Ms. Adkins, suggesting that the intrusions ended without human intervention.
Google said it first learned of the episode in July, then launched an investigation into the incident.
The debate around AI has heated up since revelations about models that have spontaneously left their confined environments to enter sites and platforms, raising fears of losing control over these tools.
Several tech leaders are advocating slowing the race to AI and a form of self-regulation within the sector.
The episodes involving Gemini “underscore the importance of training powerful AI models to act responsibly,” emphasized Heather Adkins.
