美國政府一紙命令,逼 Anthropic 全球下架 Fable 5 和 Mythos 5

美國政府以國安為由發出出口管制指令,要求停止所有外國國民——連 Anthropic 自家外籍員工都算——存取 Fable 5 和 Mythos 5。淨效果是 Anthropic 必須對全球所有客戶關掉這兩個模型。理由是政府相信有人找到了 Fable 5 的「越獄」,而 Anthropic 看完示範後說:那只是叫模型讀一個 codebase 然後修 bug——別的公開模型也做得到,這根本不算繞過。

Agent 要上 production,為什麼最後都走 MCP——Anthropic 把 API / CLI / MCP 三條路一次講透

Anthropic 剛丟出一份 production agent 的連線路指南:agent 上雲那天,API / CLI / MCP 三條路都會 ship,但只有 MCP 是會複利的那條。文章以 Cloudflare 的 MCP server(兩個 tool 包 2,500 個 endpoint)當設計 benchmark,帶出 remote 優先、tool 以 intent 分組、大 API 走 code orchestration、Elicitation + CIMD 把 production auth 收斂的整套心法。

Claude Code 四月變笨的真相——Anthropic 把三個 bug 全招了

Anthropic 剛發 postmortem 承認 Claude Code 最近一個月確實變笨——而且不是一個 bug,是三個疊在一起看起來像大型 regression。三月把 default reasoning effort 從 high 降到 medium 省 latency,一個 cache 優化的實作錯誤每 turn 都清 thinking history,四月為了壓 Opus 4.7 冗長加一段 system prompt 把 eval 打掉 3%——四月 20 全部修完,還 reset 了所有 subscriber 的 usage limit。

Opus 4.7 上路之後 prompt 該怎麼改——Anthropic 官方兩份 best practice 濃縮一次讀

Anthropic 同時釋出兩份 Opus 4.7 best practice——Claude Code 專用加完整 prompting guide。Opus 4.7 是現行最強 GA,Sonnet/Haiku 的 prompt 調教心得開始過期。這篇把兩份濃縮成一頁 cheat sheet:三件必知大事、effort 階梯怎麼選、4.6→4.7 行為差異、可直接 copy 的 prompt snippets。

Anthropic 的秘密武器:Claude Mythos Preview — 強到不敢放出來的 AI

Anthropic 發布了 Claude Mythos Preview 的 System Card — 一個強到自己都怕的 frontier model。能自主發現零日漏洞、在 Firefox 裡寫出完整 exploit,但偶爾會偷偷繞過安全限制還試圖掩蓋痕跡。這份 244 頁的報告揭開了 AI 對齊研究最前線的真實面貌。