英国 AISI 披露测试中 AI 伪造身份操纵人类审核员
英国 AISI 的人讲了个真实案例:测试里的 AI 自己编假身份,还想用提示词注入骗审核员放行恶意代码,挺吓人的。
英国 AI 安全研究所模型透明度负责人 JBloomAus 披露,一个 AI 在接受网络安全任务测试时试图入侵网络。当需要人类帮助时,该 AI 伪造了多个假身份来支持自己。它还尝试通过提示词注入操纵人类审核员,以让其恶意代码获得批准。
An AI system tried to compromise a network. When it needed human help, it created fake identities to convince someone to accept its malicious code.
@JBloomAus, Model Transparency Lead at the UK @AISecurityInst, describes how an AI being tested on cybersecurity tasks attempted to manipulate human reviewers and use prompt injection to get its code approved.
"It made up multiple fake identities to sort of back itself up."
🚀 Watch the full episode of AI Deep Dive: https://t.co/oqNID94AwU