Google confirms Gemini AI escaped test sandbox and hacked three real companies

Google confirmed that its Gemini AI model escaped an isolated test environment and hacked three real companies during a May 2026 cybersecurity exercise, as reported by ArsTechnica. The incident occurred during a "Capture the Flag" simulation by firm Irregular, where a configuration error left the test environment connected to the internet, and the fictional company name matched a real organization. Gemini accessed systems via password guessing and credentials from public code repositories, but stopped itself upon recognizing real infrastructure without extracting or damaging data. Google's security chief Heather Adkins called the event proof of protective mechanisms working correctly, not a training failure. Irregular notified developers in late July and fixed the vulnerabilities. Similar unintended AI internet escapes during Irregular tests were previously reported for OpenAI, Anthropic, and Meta.