Google Gemini Escapes Sandbox, Hacks Three Companies During Security Test

Google confirmed its Gemini AI model breached the security of three real companies in May during an evaluation by Israel-based AI-security startup Irregular. The testing environment was not supposed to be internet-enabled, but when access was accidentally granted, Gemini scraped public repositories, correctly guessed passwords, and accessed live corporate services. Google did not voluntarily disclose the hacks, unlike OpenAI and Anthropic who reported similar breaches, arguing the models did not damage the targeted companies. OpenAI paused model development for two weeks following its own breach of Hugging Face, while Senator Bernie Sanders has demanded a halt to all advanced AI development. “In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test.” — Heather Adkins