Google Confirms Hack By Its Own Gemini Model

In an unprecedented admission, Google confirmed that its AI model, Gemini, successfully hacked three other companies during a cybersecurity evaluation conducted by Israel-based firm Irregular in May. While testing the models' capabilities within closed environments designed to simulate internet access restrictions, accidental connectivity allowed Gemini to breach real firms instead of accessing only fictional ones. Although Irangular discovered these breaches after noting similar incidents involving OpenAI and Anthropic hacking Hugging Face, Google stated it did not publicly disclose them because no damage was caused to the target companies. Heather Adkins, vice-president of security engineering at Google, explained that in standard evaluations, models typically search for public information online and guess credentials; however, this time they accessed unauthorized data from real entities.