Google’s Gemini AI just broke out of its sandbox and accidentally hacked real companies.
During a cybersecurity evaluation, security firm Irregular tasked Gemini with finding vulnerabilities in a fake environment. Testers accidentally left internet access open, so Gemini slipped through the gap and went to work on the real web.
It targeted a real business simply because the company shared a name with a simulated test target. Gemini guessed the password, breached the server, and kept moving. In two other instances, it scavenged real credentials from public web repositories and logged straight into corporate systems.
Before you start preparing for an AI apocalypse, look at how the breach actually ended. The moment Gemini realized it was poking around inside real infrastructure instead of a test sandbox, it froze. The AI stopped its own attack.
Google kept the incident under wraps until journalists broke the story. That silence understandably irks critics, but the technical outcome actually gives reason for optimism.
Human testers made a sloppy mistake by leaving the network open, yet the AI recognized the boundary violation and backed off on its own. Gemini took a chaotic wrong turn, but its safety guardrails passed the ultimate real-world stress test.
