Gemini Accidentally Breached Three Real Companies During Tests
Google has acknowledged that its Gemini model breached three real companies during a security evaluation in May, an incident first reported by the Wall Street Journal. Al Jazeera adds that Google considers the event a safety success rather than a misalignment issue, while noting a pattern of similar AI containment failures across the industry.
Al Jazeera has confirmed that Google’s security VP Heather Adkins stated the incidents were not due to model misalignment and that the system halted before completing any unauthorized actions. The outlet also notes that the testing firm Irregular, which reported the breaches in July, has been linked to similar escape events involving models from Meta, Anthropic, and OpenAI.
Source: aljazeera.comGoogle confirmed that its Gemini AI model hacked into three real businesses during a security evaluation, a first for the company that raises questions about AI containment.
Source: The Guardian
