初创公司Abliteration.ai正在将移除AI模型安全防护的技术商业化。1该公司于去年晚些时候成立，今年3月正式注册，通过托管平台向用户提供已移除防护栏的开源AI模型（如GLM-5.3）的免费网络访问和API接口服务。1公司声称其目的是"支持进行攻击性网络安全、红队测试和其他模型拒绝执行的工作"。1
这一商业模式引发了安全隐忧。1测试表明，该平台可以快速生成用于窃取Chrome密码的Python程序和危险病原体培养协议等有害内容。1AI安全非营利组织CivAI研究负责人Andrew Yoon警告："abliterated模型允许你修改模型使其成为反社会者"。1Yoon还建议政府要求GPU提供商验证客户身份并阻止可疑的危险滥用。1
Abliteration.ai目前尚未融资但正在进行融资谈判。1公司已经吸引了来自英国和欧洲的多家早期红队测试初创公司作为客户。1
Abliteration.ai, a startup founded in late last year and formally registered in March, has launched a managed platform that provides access to open-source AI models with their safety guardrails removed.1 The company offers both free web access and API interfaces to models such as GLM-5.3, which have been modified to bypass standard protective measures.1
The company frames its mission around supporting red-team testing and cybersecurity defense work, stating that its goal is to "enable offensive cybersecurity, red-team testing, and other work that models normally refuse to do."1 However, testing has demonstrated that the platform can rapidly generate harmful outputs, including Python code designed to steal Chrome passwords and protocols for cultivating dangerous pathogens.1 Andrew Yoon, research lead at AI safety nonprofit CivAI, has raised significant concerns about the technology, warning that "abliterated models allow you to modify a model to make it antisocial."1
The startup, which has not yet secured funding but is currently in fundraising discussions, already counts early-stage red-teaming startups from the United Kingdom and Europe among its customers.1 Yoon has recommended that governments require GPU providers to verify customer identities and prevent suspicious dangerous misuse.1
评论
还没有评论，欢迎留下第一条。