AI Safety Requires Tracing Back Three Generations? Anthropic's Nature Study Reveals Subconscious Contagion in Large Models
36Kr
新智元
AI安全得查祖宗三代？Anthropic登Nature揭秘大模型潜意识传染·2026年04月16日 19:34AI模型只看了一串纯数字序列，就能继承另一个模型的危险偏好，即使删掉敏感词没有用，合成数据时代最隐蔽的安全裂缝，被撕开了。 刚刚，Anthropic一篇论文登上Nature，曝出了一个让整个AI安全圈坐不住的发现： 一个「坏」模型随手写的一串数字，就能「带坏」下一个模型，而且你根本看不出
const options = {
commonOptions: {
aid: 331981,
iid: '0',
did: '0',
},
captchaOptions: {
showMode: 'mask',
host: 'https://netsec-img-req-cn.zijieapi.com',
fp: 'MzQuOTYuNTEuODd8MzExMDkzMg==',
successCb () {
console.log('验证成功');
setTimeout(function() { window.location.reload(); }, 4000)
},
}
}
window.TTGCaptcha.init(options)
window.TTGCaptcha.render({
verify_data: '{"code":"10000","from":"verify_center","type":"verify","version":"1","region":"cn","subtype":"slide","detail":"CVmFZYW0SZY0pEvuTLSnJuY4kGmyQwsy7Xae-K*Ti7o5R6mXyRW2qu5V8a2QOoIeOJcbh18Mb-wiJqHAmA2f3-V7GUNaIYjnPYlbcYa1pQgo23sKP7l4TXHesiiL9KaOag*JNh18syY4USYLhAMj3dfQymBjoaazuRZ5fHR77iEaFFMp76DrUV450IerYpy4GXttbVVLsUGxCbmSAC5U7MCiGC8Q41KsCDUtHU*o5wWYiJ9N6E2opRJdRB0krpVomA9IQpHaDZJwkS5-CGf6Z5M4xWt8*FeseGz2awBk*60vO3acF6dSLW6rqN605DaYGp3peTklKwqCSGFVG*WY0g2ag8WDRhq8npLpXJPoU1MX5V80W-DXd0adzcwJKUHV484gMDekuc-jrzCmVyggwDZ4-bQgucNRqCsts86-XKPLRrtb*4i4JVm1mq9p7gd7vVuF*7Khj--KHp5T","verify_event":"","fp":"verify_a160144b17f98b38388d6dbe3cae35fc","scene":"","verify_ticket":"","channel_mobile":"","sms_content":"","mobile":"","email":"","sec_user_id":"","auth_ticket":""}',
})
window.TTGCaptcha.getFp().then((fp) => {
})