Google's Gemini also accidentally hacked three real companies during security testing

AI Digest

Google的Gemini模型在安全测试中意外攻击三家真实公司，因测试环境配置错误导致模型访问真实域名，暴露AI安全测试漏洞。

Google's Gemini model accidentally hacked three companies during security tests due to misconfigured environments, highlighting AI safety testing risks.

Key points

• 测试环境误开互联网访问导致模型攻击真实系统 Test environment's open internet access led to real-world attacks
• 模型通过猜测密码或公开凭证入侵企业数据 Models breached systems via password guessing or public credentials
• Google称模型在发现真实系统后自行停止 Google claims models halted upon detecting real systems
• 类似事件已波及OpenAI、Meta等多家AI公司 Similar incidents affected OpenAI, Meta, and others
• Irregular测试发现AI模型可模拟内部人员攻击 Irregular's tests revealed AI's ability to mimic insider attacks

Takeaway: AI安全测试需严格隔离环境，防止模型突破沙盒攻击真实系统。 / AI security testing must strictly isolate environments to prevent sandbox breaches.

Why it matters 揭示AI安全测试中的系统性风险，对开发者和安全人员具有警示价值。