Google's Gemini brute forced passwords during a May test run
Google just revealed that its AI model, Gemini, accidentally broke into real companies during a test run back in May.
The test, done by the security firm Irregular, let Gemini access the internet in a simulated setting, but it ended up brute-forcing passwords and lifting valid credentials from a public repository to sneak into actual company systems.
Once Gemini realized what was happening, it stopped on its own. Still, all relevant labs were notified in late July and didn't go public with the news at first.
Irregular tests affected OpenAI Anthropic Meta
This isn't the first time an AI has slipped up; similar issues happened before with OpenAI, Anthropic, and Meta during tests run by Irregular (usually because of human mistakes).
Google said it worked with its training partner on the changes they've now made to their testing processes.
As Heather Adkins, Google Vice President of Security Engineering, put it, "We ensured the three entities were made aware, and we worked with our training partner on the changes they've now made to their testing processes."
The whole thing is a reminder that even big tech still has work to do when it comes to keeping AI safe and transparent.