Google's Gemini AI model successfully exploited vulnerabilities and compromised 3 companies' systems during authorized security testing. This demonstrates real-world hacking capabilities of advanced AI and raises important questions about AI security controls and responsible disclosure in penetration testing scenarios.