跟讀練習: Tragic mistake... Anthropic leaks Claude’s source code - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

78 句

如果句子過短或過長，請點選 Edit 進行調整。

Yesterday, the most ironic thing ever happened.

Anthropic, a $380 billion startup built on the idea of safety first, that advocates for closed-source software for the supposed benefit of humanity,

a company Elon calls Miss Anthropic, whose logo is definitely not a sphincter, whose CEO has been warning us for years that human programmers will be replaced by AI in six months,

just accidentally leaked Claude Code's entire source code to the internet at 4am, officially making Anthropic more open than OpenAI.

Within minutes of the leak, Chaofan Xiao, a security researcher,

discovered that version 2.1.88 of the Cloud Code NPM package was shipped with a 57MB source map file.

You know, that file that's only used in development with the full readable source code of your project.

This holy grail of leaks containing over 500,000 lines of TypeScript code quickly spread across the internet like wildfire.

Anthropic's legal team courageously issued DMCA takedowns, but by the time they woke up in San Francisco, it was already too late.

The code was mirrored countless times and cloned by slop tubers like Fireship, which the Supreme Court says I can do legally, by the way, as a world-renowned journalist.

I know a lot about the law and various other lawyerings, um...

In today's video, we'll look at all of the incredible discoveries in the code that Anthropic doesn't want you to know about.

Like its anti-distillation poison pills, its mysterious unreleased features, its undercover mode, its regular expression-based frustration detector,

and many other super-secret techniques at the foreskin of AI research.

No, this is not an April Fool's Day joke, it is April 1st, 2026, and you're watching the code report.

Unfortunately, my lawyer just informed me that showing you Anthropik's TypeScript code would be a violation of my parole, and I refuse to go back to jail.

But luckily, the open source community has already created a loophole.

Ironically, Claude's most prolific user used OpenAI Codex to rewrite Claude Code's TypeScript code to Python code, resulting in a new, barely legal project called ClawCode,

and it's already become the fastest repo in history to surpass 50,000 GitHub stars.

Not only that, but somebody else slop-forked the leaked code and made it work with any model.

They're calling this new project OpenClaude, and it makes projects like OpenCode completely obsolete.

But maybe, now that the code's out in the open, Anthropic will just make it open source.

Somebody tried to make a pull request with the leaked code, but not surprisingly, it looks like Anthropic already deleted it.

What a mother-effin' crazy 24 hours, but how did this code end up leaked in the first place?

Well, as I mentioned, the source map was accidentally packaged in an NPM release, but that's weird because build tools normally strip out source maps automatically.

Well, Cloud Code is built on BunJS, which as you might recall was recently acquired by Anthropic.

And it just so happens that about three weeks ago, somebody opened up an issue on GitHub about BunJS serving source maps in production.

Wouldn't it be ironic

if the fastest JavaScript runtime in the world also turned out

to be the fastest way to ship your entire codebase to the internet?

It's unclear if that was the root cause, and it's also possible that some unfortunate developer did this by accident, or perhaps some rogue developer did it on purpose.

We may never know the truth, but now it's time for the fun part.

What did we actually learn from the leak?

Well first, we learned that Claude uses Axios.

If you subscribe to my channel, then you know that Axios was compromised by North Korean hackers yesterday.

The exploit can install a remote access Trojan on anyone using this package, and in theory, if that happened on Anthropic servers, it could be a massive disaster.

But the next thing we learned is that Claude code is basically just a dynamic prompt sandwich glued together with TypeScript, and not some magical piece of futuristic technology.

In a basic AI chatbot, you typically have a hidden system prompt that gets combined with your prompt,

then the base model uses statistics to regurgitate a bunch of data it stole from the internet.

But in Claude code, things are far more complex, with a total of 11 steps from input to output.

Somebody already vibe-coded a website that breaks down every step, but the most interesting part about this codebase is

that it contains tons of hard-coded instructions and guardrails that basically beg Claude to please don't do anything weird.

Like there's just file after file of these massive hard-coded strings telling Claude to be a good boy.

And that's kind of surprising, because if this code were ever leaked, it would instantly turn from a black box into a blueprint for Claude's competition.

And ironically, that's exactly what happened.

What makes that even more funny, though, is that Claude was actively trying to stop their competition from copying Claude code by implementing anti-distillation poison pills.

It does that by pretending that certain tools exist, when in reality, they don't exist at all.

You're a big fat That means if you're some Chinese guy trying to train a new model on Claude's outputs, it's going to talk about tools that don't exist,

which will point your model in the wrong direction and just make it suck.

In reality, Claude Code only uses about 25 different tools or so, and now the Claude distillers know exactly what to look for, and they're likely going to have a field day with the Bash tool.

This file contains over a thousand lines of code that helps the large language model reliably parse and execute Bash commands,

which might be the single most important feature in an AI coding assistant.

The next thing we need to talk about, though, is undercover mode, which is a set of instructions that tell Claude to never mention itself in commit messages or outputs,

where the main idea is to make the outputs look as human as possible.

The stated purpose of this feature is to prevent things like model codename leaks, but many have speculated that the true purpose is more deceptive,

like they're trying to covertly use Claude in open-source projects so AI code doesn't get scrutinized when it breaks things catastrophically.

A very misanthropic idea, indeed.

But the irony continues.

Another funny thing found in the code is its regex frustration detector.

Your state-of-the-art AI model uses simple regular expression matching against your prompt to look for keywords like shit, fuck, cock, balls, and so on,

to determine if you're not having a good experience coding with Claude.

If it detects a match, it'll simply log an event.

The bottom line here is that we're not looking at some sort of alien superintelligence, but rather basic programming concepts that have been around for 50 years combined with a bunch of prompt spaghetti.

It's all just an illusion.

On top of that, this codebase has a ton of comments.

A lot more comments than you would typically find in a human-written codebase.

And what that tells us is that these comments aren't actually written for humans, but rather for the AI to write its own AI coding tool in an infinite loop.

But perhaps the biggest problem about this League 4 Anthropic is not the code itself, but rather the feature names and roadmap hidden within the code.

Like there's a hidden capability under a feature flag called Buddy,

which appears to be a new Tamagotchi-style companion that every developer can customize and raise like a little digital pet.

This might just be Anthropik's April Fool's Day joke, but there are also references to Opus 4.7 and a new model called Capybara,

which might be their new recently-teased Mythos model.

There's also things like Ultra Plan, Coordinator Mode, and Demon Mode, but perhaps the most interesting is Kairos,

which is a Greek word for an exact moment in time, or God's time.

I hate to beat off a dead horse here, but it's a bit ironic that Anthropic didn't get to reveal Kairos at the exact time it wanted to, and instead, God chose the right time.

The feature itself seems to be some kind of background agent that keeps a daily journal, uses dream mode to consolidate memories, and does work for you in the background on a specific schedule.

Pretty cool, but at the end of the day, this leak is a pretty huge setback for Anthropic, which hopes to IPO later this year and offload their bags to the retail public.

And it's yet another reminder that your top secret application is just one NPM publish away from becoming open source, whether you like it or not.

下載應用

流利口語所需的一切

AI 發音評分逐句打分

音標練習掌握每個音

詞彙學習建立你的詞庫

詞彙遊戲邊玩邊學

掃碼

Download on theApp Store Get it onGoogle Play

同頻道

FOR YOU

本課的詞彙與口語要點

這節 C1 級別的口語課以影片“Tragic mistake... Anthropic leaks Claude’s source code”為素材。影片中反覆出現的詞有：code, Claude, Anthropic, source, model。這段影片共有 78 個句子、1477 個單詞可供跟讀。講話部分時長為 7:22。說話人語速較快，每分鐘約 200 個詞，會出現較多連讀和弱讀。只有 78% 的單詞屬於英語最常用的 3,000 詞，詞彙難度較高。

影片中的重點詞彙

影片中最難的 15 個單詞，附發音和釋義：

單詞	發音	釋義
leak 名詞	/liːk/	漏洞
codebase 名詞		程式碼庫 /程式碼庫
distillation 名詞	/dɪstɪˈleɪʃən/	蒸餾 /蒸餾
detector 名詞	/dɪˈtɛktə(ɹ)/	檢測儀 /檢測儀
frustration 名詞	/fɹʌsˈtɹeɪ.ʃən/	挫折
pill 名詞	/pɪl/	藥丸 /藥丸, 藥片 /藥片
poison 名詞	/ˈpɔɪ.zən/	毒, 毒物
blueprint 名詞	/ˈbluːˌpɹɪnt/	藍圖 /藍圖
chatbot 名詞	/ˈtʃætbɑt/	聊天機器人 /聊天機器人
consolidate 動詞	/kənˈsɒl.ɪ.deɪt/	聯合 /聯合, 合併 /合併
customize 動詞	/ˈkʌs.tə.maɪ̯z/	定製, 定做
deceptive 形容詞	/dɪˈsɛp.tɪv/	迷惑的
guardrail 名詞	/ˈɡɑɹdˌɹeɪl/	護欄 /護欄
keyword 名詞	/ˈkiː.wɜːd/	關鍵詞 /關鍵詞, 關鍵字 /關鍵字
loophole 名詞	/ˈluphoʊl/	漏洞

影片中出現的短語動詞

單詞	釋義
break down 動詞	壞了 /壞了, 出故障
end up 動詞	到達
turn out 動詞	原來 /原來

影片中的語法

說話人最常用的結構，並附上影片中的原話：

結構	影片中的用法
條件句 if + 從句，will/would + 動詞 — 條件及其結果	If it detects a match, it'll simply
被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的	will be replaced · was shipped · was mirrored
現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事	has already created · have speculated · have been

需要注意的發音

說話人用了 18 次縮略和弱讀形式，例如 you're, don't, they're。請按聽到的簡短形式來說。

“sh” 和 “zh” 音: distillation /dɪstɪˈleɪʃən/, frustration /fɹʌsˈtɹeɪ.ʃən/, illusion /ɪˈl(j)uːʒ(ə)n/
長單詞——注意重音位置: distillation /dɪstɪˈleɪʃən/, ironically /aɪˈɹɑnɪkli/, undercover /ˌʌndə(ɹ)ˈkʌvə(ɹ)/, consolidate /kənˈsɒl.ɪ.deɪt/, futuristic /ˌfjut͡ʃəˈɹɪstɪk/

如何用這段影片練習

先完整聽一遍影片，不要開口，記下不認識的單詞。
先用 0.75 倍速逐句跟讀，熟練之後再回到正常速度。
錄下自己的聲音並與原聲對比，特別注意 leak, codebase, distillation 這類單詞。

什麼是跟讀法？

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →

跟讀練習: Tragic mistake... Anthropic leaks Claude’s source code - 通過影片學習英語口語

選擇練習模式

下載應用

The last real programmers are dying out...

PostgreSQL in 100 Seconds

React Native in 100 Seconds

Did Google just dethrone OpenAI and Anthropic?

A $6.3 billion open-weight model just got embarrassed by the French...

Every app on Earth is now open source... whether it likes it or not

OpenAI’s lead engineer was secretly a spy...

Mind-bending new programming language for GPUs just dropped...

Intel chips can’t possibly be this bad… 100% crash rate?

Godot in 100 Seconds

Grok 4 pushes humanity closer to AGI… but there’s a problem

Array Map in 100 Seconds

Docker in 100 Seconds

OpenAI’s new browser feels familiar…

Angular in 100 Seconds

10 weird OSS projects you need right now...

Firefox and Tor hit with 9.8 critical level exploit

5 open source tools that replaced my $320/mo AI stack...

The most interesting "hack" in history...

C++ in 100 Seconds

When open-sourcing your code goes wrong...

The mystery is solved... and the answer is 40x cheaper than Claude

Claude Code: The Advanced Guide (99% of Devs Skip These Features)

Last Cheater Standing Wins $10,000!

Oh, wait, actually the best Wordle opener is not “crane”…

케빈 시스트롬: 행동의 열매

AI is flooding the internet—here’s what I’m protecting

Did Anthropic Kill Git?

Product Recall: A Crisis Call with the PR Team

What is clean code? - Uncle Bob

He’s always wrong

Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC Berkeley