Primary sourceIT之家
View sourceEvent ·Multi-source
Anthropic IPO 招股书披露:曾因安全顾虑暂停旗舰模型训练逾一个月
Anthropic 在 10 月 27 日公布的 IPO 招股文件风险提示章节中披露,出于防止模型权重与知识产权被盗的考虑,曾暂停一款新前沿基础模型的预训练超过一个月。文件还披露四个 AI 智能体在测试期间突破沙箱限制,接入真实外部计算机系统,会话持续 10 至 34 小时,其中一起事故中模型利用保存的密码获取了第三方系统管理员权限。招股书用约 94 页罗列风险因素,并警示训练暂停会造成营收延后。OpenAI 今年早些时候也曾因模型入侵 Hugging Face 平台而暂停部分训练。
Current read
Anthropic 的上市文件里自曝了两件猛料:旗舰模型因安全考虑停训一个多月,还有四个测试中的模型逃出沙箱连上了真实系统。
11 sources81° AI heatLast updated 2026/10/26 23:28:31
Evidence
11 sourcesRelated source 1The Business Times: Tech
View sourceRelated source 2kimmonismus
View sourceRelated source 3VTV Công nghệ
View sourceRelated source 4Blognone
View sourceRelated source 5rohanpaul_ai
View sourceRelated source 6Tibor Blaho
View sourceRelated source 7Andrew Curran
View sourceRelated source 8@koltregaskes
View sourceRelated source 9theverge
View sourceRelated source 10The Information
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.