Anthropic模型在受控安全测试中成功入侵三个组织
人工智能开发商Anthropic证实,其模型在内部测试过程中被用于渗透三个不同的组织。

Anthropic近期进行了安全评估,其人工智能模型被指派尝试入侵三个不同的组织。这些测试旨在探索先进AI系统在与外部数字基础设施交互时所具备的能力及潜在风险。
据《The Edge Malaysia》报道,在这些受控演习中,AI模型成功获得了对三个目标实体的未经授权访问权限。报告指出,这些实验是更广泛研究工作的一部分,旨在更好地理解在现实场景中部署强大的生成式AI工具所带来的安全影响。
这些测试结果突显了AI安全性不断演变的特性,以及这些系统以绕过标准数字防御方式被利用的可能性。Anthropic已利用这些发现来完善其模型的安全协议,旨在防止此类漏洞在受控研究环境之外被恶意行为者利用。
对于马来西亚的企业和技术领导者而言,这些进展强调了加强网络安全框架以应对新兴AI驱动威胁的紧迫性。随着本地企业越来越多地将生成式AI整合到运营工作流程中,这些测试中所识别出的风险,对于那些需要在技术创新与维护数字资产安全之间取得平衡的企业而言,是一个关键的提醒。确保AI工具配备充足的保障措施,正日益成为企业的核心优先事项。
Source
Originally reported by The Edge Malaysia. Read the original report →
This story was translated from our English report. Read in English →
Join the conversation
We post stories like this all day on Threads. Discuss this story on Threads →
More in AI
Google Integrates Gemini AI To Enhance Robotic Dexterity
New developments from Google aim to improve how robots handle complex physical tasks using advanced artificial intelligence.

Amazon Posts Strong Q2 Growth Driven by AI and Cloud Expansion
Amazon shares surged after reporting a 20% revenue increase as its cloud and artificial intelligence divisions exceeded performance expectations.

OpenAI CEO Sam Altman to Engage Trump Officials on AI Safety
Sam Altman is set to discuss voluntary safety protocols with the incoming administration following reports of an AI agent operating outside its parameters.

Goldman Sachs Asset Management Launches Dedicated AI Investment Platform
The financial giant has moved to consolidate its artificial intelligence strategy through a new dedicated investment division.
