可能窃取你全部积蓄的新AI模型

NPR。

NPR.

四月份，财政部长斯科特·贝森特和美联储主席杰罗姆·鲍威尔召集了一次银行首席执行官的特别会议。

In April, treasury secretary Scott Bessent and Federal Reserve Chair Jerome Powell called a special meeting with bank CEOs.

他们想确保在人工智能快速发展的时代，银行的数字金库依然安全。

They wanted to make sure banks were keeping their digital vaults secure in an age of rapidly advancing AI.

这是对人工智能公司Anthropic的一则公告作出的回应。

They were responding to the AI company Anthropic's announcement.

这家人工智能公司表示，他们开发出了一款全新的模型，技术先进、极其聪明，但也非常危险，因此暂时无法向公众发布。

The AI company said it had a new model that was so state of the art, so clever, so dangerous that it couldn't yet release it to the wider public.

Anthropic把这款模型命名为Claude Mythos。

Anthropic called the model Claude Mythos.

迈克尔·摩尔负责该公司网络安全产品。

Michael Moore heads up the company's cybersecurity products.

他说，这个新模型能够识别成千上万的软件漏洞。

He says the new model could identify thousands of software vulnerabilities.

这意味着，当初写这段代码的人，某个地方没处理好。

What that means is someone who was writing the software at some point didn't quite do something right.

结果就留下了一道没关严的门。

And that left a door kind of ajar.

攻击者就能从这道门溜进来。

And an attacker can get in through that door.

这是《Planet Money》的《指标》节目。

This is The Indicator from Planet Money.

我是达里安·伍兹。

I'm Darian Woods.

我是韦隆·王。

And I'm Waylon Wong.

今天节目中，我们的钱在银行里还安全吗？

Today on the show, is our money still safe in our bank accounts?

我们会探讨越来越强大的人工智能这把双刃剑，也会问一问，Claude Mythos 的独特能力，到底是真本事，还是虚张声势。

We learn about the double edged sword of increasingly capable AI, and we ask whether Claude Mythos' unique capabilities are just a myth or a meef.

本节目由MS Now出品。

This message comes from MS Now.

人工智能的发展速度真快。

AI is moving fast.

哪些是真实的？

What's real?

哪些是炒作？

What's hype?

这一切最终会走向哪里？

And where is it all headed?

关于为什么会发生这些，克里斯·海斯采访了多位顶尖专家，帮我们理清头绪。

On why is this happening, Chris Hayes talks with the leading experts to make sense of it.

《人工智能的终局》，MS Now特别系列。

The AI endgame, a special series from MS Now.

欢迎在您收听播客的平台收听。

Listen wherever you get your podcasts.

新的AI模型Claude Mythos，在某些方面跟普通的语言模型没什么两样。

The new AI model, Claude Mythos, is, in some ways, a typical language model.

它跟Anthropic其他Claude模型或者ChatGPT一样，主要就是用来回答问题或者写代码，并没有本质上的不同。

Wasn't fundamentally different from a chatbot change to answer general questions or write lines of code like Anthropic's other Claude models or ChatGPT.

它并没有专门针对网络安全进行过训练。

It hasn't been trained specifically for cybersecurity.

所以它能写诗、写打油诗，甚至给你推荐一款顶级的巧克力曲奇食谱。

So it could write poetry, limericks, or give you a new recipe for the best chocolate chip cookies.

但当Anthropic测试Claude Mythos时，公司里的一些人，比如迈克尔·摩尔，开始感到不安。

But when Anthropic was testing Claude Mythos, people in the company like Michael Moore started getting concerned.

它展现出了一种非常高级的能力，能把多个独立的漏洞串联起来，形成一个可能极具破坏性的攻击手段。

It showed a a really advanced capability of chaining together individual vulnerabilities into an exploit that was potentially really dangerous.

迈克尔说的‘串联漏洞’，意思是Mythos能发现一些单独看问题不大的代码片段，但合在一起就会让软件出现严重漏洞。

When Michael says chaining together individual vulnerabilities, he means that Mythos could find multiple lines of code that on their own weren't hugely problematic, but together could leave the software exposed.

这就像一个小偷，先撬开公寓楼的大门，再打开大楼的正门，最后还能打开你家的门锁。

It's like if a thief could break into your apartment block's front gate and your building's front door and your apartment unit's lock.

这是我们第一次意识到，如果直接把模型公开发布，AI的规模可能让攻击者以前所未有的速度入侵系统，这让我们非常担忧。

It was the first time that if we just were to release the model out into the public, we were concerned that the scale of AI would allow offensive actors to actually go break into systems at an unprecedented rate.

因此，Anthropic 只给少数几家公司开放了对 Mythos 的特殊访问权限。

So Anthropic gave special access to Mythos to only a small group of companies.

目的是让他们在模型全面发布前，先修复自己软件中的漏洞。

This was to allow them to fix their software before Mythos had a wider release.

Anthropic 把这个项目叫做‘Glasswing’。

Anthropic called it Project Glasswing.

我们希望给这些公司争取足够的时间，避免大规模的安全事件波及所有人。

We wanted to make sure that we gave a good head start to, avoid a a large swath of incidents, that could impact all of us.

最初，只有一家金融机构获得了早期访问权限，那就是摩根大通。

At first, there was only one named financial institution given early access, JPMorgan Chase.

随后，《路透社》报道，美国银行是第二家。

Reuters then reported that Bank of America was two.

也就是说，在美国四千多家银行里，只有两家得到了访问资格。

That's two out of America's 4,000 banks.

人们该担心自己的积蓄会从银行账户里凭空消失吗？

Should people be worried about their life savings potentially disappearing from their bank accounts?

我很高兴看到大家正在积极讨论，也在努力修补尽可能多的安全漏洞，避免让大多数人陷入严重担忧的境地。

I'm really excited to see the conversations that are happening and the work that's being done to patch as many vulnerabilities as possible so that it doesn't come to a a moment of of strong concern for most people.

希望这件事最后能像千禧年问题那样，其实没造成什么大影响。

Hopefully, it's a little bit kind of like what y two k turned out to be, which was really not much.

但那是因为当时投入了大量工作，才确保了问题没有发生。

But that was because a lot of work went into making sure it didn't happen.

我很欣慰看到整个行业现在正在做类似的努力。

And I I'm really excited to see the industry making those kind of moves right now.

韦伦，1999年的时候，你担心过千年虫问题吗？

Were you worried in 1999 about the y two k bug, Waylon?

我当时更担心的是怎么好好庆祝1999年，玛丽。

I was more concerned with partying like it was 1999, Mary.

那时候的优先事项就是这样。

Those were the priorities.

当然了，由于全球各地的程序员和技术人员都在努力工作，最后我们其实也没必要太担心。

And, of course, with the computer programmers and technicians working all over the globe, we didn't actually have to worry about much in the end.

对。

Yes.

但他们真能对像Mythos这样的AI模型也这么做吗？

But can they really do the same thing with their AI models like Mythos?

这听起来难度很大，毕竟全球有那么多公司、银行，还有各种机构都在使用软件。

That sounds like a very challenging proposition because there are so many, I don't know, companies and banks and all kinds of people who use software all around the world.

那么，像Anthropic这样的公司，到底该怎么提前向全世界发出警告，防止模型落入错误的人手中呢？

So how logistically can a company like Anthropic warn the world ahead of the model getting into the wrong hands?

是的。

Yeah.

其实我们能做的主要有两件事。

So there's really two things that we can do.

我们希望公开讨论这个模型的强大能力，以及它带来的安全风险。

We wanted to talk publicly about the power of the model and the safety concerns of the model.

但我们第二件事是确保尽可能多的人能快速修补这些安全漏洞。

But the second thing we wanted to do was make sure that there was a good quick path for as many people to patch vulnerabilities as possible.

Mythos 能把多个漏洞串联起来使用。

Mythos can chain together a lot of vulnerabilities.

你可以把它想象成一条真正的链条。

You think of it literally like a chain.

只要打断链条中的任意一环，整个攻击就失效了。

If you break just one link in that chain, the exploit doesn't work anymore.

对吧？

Right?

正是出于帮助全球软件修复漏洞的初衷，Anthropic 才推出了下一个产品。

This drive to help patch up the world's software is partly why Anthropic released its next product.

四月下旬，Claude 推出了一款专门用于发现和修复软件漏洞的工具。

Later in April, Claude released a special tool just for finding and fixing software vulnerabilities.

这款工具叫 Claude Security。

It's called Claude Security.

目前，它并没有基于那个前沿的Claude Mythos模型。

At this moment, it's not based on that cutting edge Claude Mythos model.

但迈克尔指出，即使使用稍逊一筹的AI模型，也能发现大量漏洞，然后加以修复。

But Michael points out that even using slightly less state of the art AI models can still find a lot of bugs which can then be fixed.

这听起来有道理，但一个持怀疑态度的人可能会说，这时间点也太巧了。

This makes sense, but a skeptic might say this is convenient timing.

你怎么回应那些认为Anthropic在制造问题的批评者？

What do you say to critics who might argue that Anthropic is creating the problem?

它是在让人花钱来解决自己制造的问题。

It's charging people money to solve.

这就像一家公司先造出一把万能钥匙，能打开任何门，然后又去卖锁。

You know, it's almost like a a company that creates this new skeleton key that can break in anywhere is now selling new locks.

这是个好问题。

It's a good it's a good question.

我觉得这里有几点需要说明。

I think there's a couple of things there.

云安全的核心，其实是审视那些已经存在了几十年的代码库。

Cloud security is really about looking at code bases that have been around for for decades.

我们在Mythos博客里提到过一个发现：在互联网基础设施的关键组件中，有一个已经存在了27年的漏洞。

One of the things that we found in and we discussed in our, Mythos blog was actually a bug that has been around for twenty seven years in a key piece of of Internet infrastructure.

这类漏洞正是我们期待AI在未来能发现的类型，所以现在用AI来修复它们显得尤为重要。

Those kind of bugs are the things that we expect AI to find in the future, and that's why patching them using AI is so important now.

随着行业不断发展，越来越多的代码由AI生成，我们完全有理由相信，整个行业的安全性会越来越好。

As we move and continue to evolve as an industry, and more code is written by AI, we definitely expect the security of the industry to get better.

我们一直在全力投入，确保我们的模型生成的代码本身是安全的。

It's something that we work on very strongly, to continue to make sure our models are writing secure code.

我相信其他同行实验室也在做同样的事。

I'm sure our other peer labs is due as well.

通过采用云安全这类最佳实践，我们能在问题进入生产环境、影响用户之前，尽可能多地发现并修复漏洞。

And by using best practices like cloud security, for instance, we can assure that we're finding all of the bugs that we can before things get into production and before things impact end users.

但其他网络安全专家并没有这么乐观。

But other cybersecurity experts aren't as confident.

蕾切尔·格林施塔特是纽约大学的计算机科学教授。

Rachel Greenstadt is a computer science professor at New York University.

她指出，经验较少的开发者正在‘凭感觉写代码’，也就是直接给聊天机器人下指令，让它帮自己写程序。

She points out that less experienced developers are vibe coding software, as in just typing instructions to get the chatbot to build the programs for them.

人们现在能更快地写出更复杂的东西，但自己根本不懂，AI也不懂。

People are able to write much more complex things, much quicker that they don't understand, that the AI doesn't understand.

这为新型漏洞的出现创造了机会。

And this is an opportunity for whole new classes of bugs.

一直以来，高级黑客攻击和网络安全工具之间都像猫鼠游戏一样相互博弈。

Now there has always been a cat and mouse game between more sophisticated hacking and tools that help with cybersecurity.

今年，美国情报界指出，来自中国、俄罗斯、伊朗、朝鲜的势力以及勒索软件团伙，依然对美国的网络和基础设施构成严重威胁。

This year, The US intelligence community noted that actors from China, Russia, Iran, North Korea, and ransomware groups continue to pose critical threats to US networks and infrastructure.

所以，这可能只是这场博弈的又一章，也可能Claude Mythos真的不一样。

So maybe this is just another chapter in that game, or maybe Claude Mythos is different.

这绝对是出色的营销。

I mean, it's excellent marketing.

对吧？

Right?

你知道，这种说法——说有个新模型特别危险，只有少数人能用，诸如此类的。

You know, this this idea that there's this new model, and it's it's so dangerous that only some people can have it and and so on.

不过，它发现的这些漏洞确实是真实存在的。

That said, the vulnerabilities it's finding are real.

但你也知道，这种事我们以前见过。

But, you know, we've seen this sort of thing before.

现在市面上已经有好几款AI模型，能力跟Claude Mythos差不多了。

There are AI models available out there that are already roughly as capable as Claude Mythos.

事实上，有一个开源模型早就发现了迈克尔提到的那个漏洞，那个藏在旧操作系统里二十多年的bug。

In fact, an open model found that bug that Michael had talked about, that twenty seven year old bug lurking in an old operating system.

OpenAI后来也公开发布了一个模型，功能上确实跟Mythos相当。

OpenAI has since released a model publicly that actually seems to match Mythos and what it can do.

我们问过Anthropic的迈克尔，关于这件事。

We asked Michael at Anthropic about that.

Anthropic 对其他公司发布类似强大工具这件事怎么看？

How is Anthropic thinking about other companies putting out similarly powerful tools?

是的。

Yeah.

我们当然没法评价那些没测试过的其他实验室的模型或相关产品。

Well, we we can't obviously speak to, models or or things that we haven't tested from other other labs.

我能说的是，我们目前在 Anthropic 做的这些努力，正在推动更多人讨论这些问题，也有很多人在主动修补漏洞。

What I can say is, again, I I am optimistic that the moves that we are making at Anthropic right now are fostering these conversations, and so many people are out there patching vulnerabilities.

我们也希望，无论用户用的是哪个模型，都能去这么做。

And we hope that users using whatever models they have access to go and do the same.

我们希望，这个推动者就是我们自己。

And we hope that that's us.

但如果不是我们，最重要的是漏洞必须被修补，真正提升网络世界的整体安全性。

But if it's not, it's most important that the vulnerabilities are patched and and meaningfully increase the security of of the cyber world.

对 Rachel 来说，我们正进入一个更快的新时代，但面临的挑战本质上并没有超出我们过去处理过的范围。

To Rachel, we are in a new faster era, but not one that brings challenges that are fundamentally different from what we've tackled before.

据我们所知，既没有天使现身，也没发生天塌下来的事。

It's neither the case that the angels have appeared nor the case that the sky is falling as far as we know.

明白了。

Okay.

听到这个真让人松口气。

That's relieving to hear.

韦隆，你感觉怎么样？

How are feeling, Waylon?

我觉得咱们该像过1999年那样狂欢一场。

I'm feeling like we should party like it's 1999.

该像过2026年那样狂欢。

Party like it's 2026.

本集由库珀·卡茨为金制作，音效由吉米·基利负责，校对由西耶拉·华雷斯完成。

This episode was produced by Cooper Cats for Kim with engineering by Jimmy Kealey, was fact checked by Sierra Juarez.

基克和卡坦负责剪辑，《指标》节目由NPR出品。

Kick and Catan edits the show and The Indicator is a production of NPR.