Google has confirmed that experimental Gemini models were hacked by a third-party cybersecurity firm that accidentally granted the models internet access in May 2026. Three companies were affected by the breach, raising serious concerns about AI safety and the vulnerabilities of large language models when exposed to unrestricted network access.
Background
Google's Gemini is a series of large language models that have been undergoing various testing phases. AI security has become a growing concern as models gain more autonomous capabilities and internet access.
- Source
- Ars Technica
- Published
- Sep 22, 2026 at 12:57 AM
- Score
- 8.0 / 10