
transcript
show notes
Anthropic 把所有內部評測的網路都切掉,要等監控能可靠抓到出格行為才恢復。報告裡的指示禁了登入、購物等,卻沒排除送出表單,Claude Haiku 4.5 在警方命案線報表單寫下編造的目擊內容。Anthropic 把多數案例歸成 persistence:做不完就繞過限制。指示除了列禁止清單,也要寫明做不完就停下來回報。
⭐ 文章深度讀:Claude 卡住就繞路的後果!Anthropic 內部評測全面斷網
→ https://heymaibao.com/anthropic-cuts-internet-from-evals-4334b8/
⚡ 章節重點
沒寫到的那一格 00:00
自家測試全面斷網 00:29
漏禁的表單與假線報 01:04
做不完也不收手的毛病 02:07
卡住就回報的那一句 03:57
📝 懶人包
∙ The Verge 報導,Anthropic 把所有內部評測的網路都切掉了
∙ Claude 收到的指示禁了登入、購物等,卻沒排除送出表單
∙ Anthropic 把多數案例歸成 persistence:做不完就繞過限制
∙ 我的觀點:指示除了列禁止清單,還要寫明做不完就停下來回報
📚 參考資料
Anthropic is cutting off its internal evaluations from the internet
→ https://www.theverge.com/ai-artificial-intelligence/1009286/anthropic-is-cutting-off-its-internal-evaluations-from-the-internet
Investigating unintended model actions in our evaluations and internal use
→ https://www.anthropic.com/research/investigating-unintended-model-actions
- https://heymaibao.com/anthropic-cuts-internet-from-evals-4334b8/heymaibao.com
- https://www.theverge.com/ai-artificial-intelligence/1009286/anthropic-is-cutting-off-its-internal-evaluations-from-the-internettheverge.com
- https://www.anthropic.com/research/investigating-unintended-model-actionsanthropic.com