等等……GPT-6 究竟有多强？

今天《AI每日简报》带来一起安全事件，让我们不禁要问：GPT-6到底有多靠谱？

Today on the AI Daily Brief, a security incident that has us asking, just how good is GPT-six really?

在那之前，先看下头条新闻：谷歌发布了新模型，但可能不是我们最期待的那些。

Before that in the headlines, a new set of Google models, but not necessarily the ones that we wanted.

《AI每日简报》是一档每天更新的播客和视频节目，聚焦AI领域最重要的新闻与讨论。

The AI daily brief is a daily podcast and video about the most important news and discussions in AI.

好了，各位。

Alright, friends.

在正式开始之前，先说几件事。

Quick announcements before we dive in.

首先，感谢今天节目的赞助商：普华永道、Rackspace、Blitzy 和 Airtable。

First of all, thank you to today's sponsors, KPMG, Rackspace, Blitzy, and Airtable.

想收听无广告版本？请访问 patreon.com/aidailybrief，或者在 Apple 播客订阅我们。

To get an ad free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts.

如果想了解如何赞助本节目，请发送邮件至 sponsors@AIdailybrief.AI。

And to learn more about sponsoring the show, send us a note at sponsors@AIdailybrief.AI.

最近大家都在讨论各种新模型，唯独谷歌显得格外沉默。

In all of the recent model talk, one lab that has been kind of conspicuously absent is Google.

已经好几个月了，谷歌都没再更新他们的Pro系列模型，我们只能继续用像3.5 Flash这样更小更快的版本。

It has now been months and months since we got any sort of update from them on their Pro series models, having to have contented ourselves with just smaller and faster models like 3.5 Flash.

昨天的发布并没有推出传闻中表现不佳的3.5 Pro。

Yesterday's announcement did not bring 3.5 Pro, which has been rumored to be underperforming.

相反，我们又看到一批新的Gemini Flash变体。

Instead, we once again got a set of new variants of Gemini Flash.

周二发布的重点是Gemini 3.6 Flash，最大的改进是提升了令牌使用效率。

Tuesday's release was headlined by Gemini 3.6 Flash, and the big change is better token efficiency.

在人工分析基准测试中，这个新模型比3.5 Flash少用了17%的令牌。

On the artificial analysis benchmark run, the model used 17% fewer tokens than 3.5 Flash.

谷歌还提到，在某些特定测试如DeepSUI上，令牌使用量最多减少了65%。

Google also said that on some isolated benchmarks like DeepSUI, they observed up to a 65% reduction in token usage.

这一点尤其重要，因为当初3.5 Flash发布时，最受诟病的就是它比前代模型贵得多，令牌消耗也大得多。

Now this might be particularly relevant because one of the loudest complaints around the release of 3.5 Flash was that the model was significantly more expensive and heavy on token usage than its predecessors.

谷歌似乎更注重速度优化，但这让人开始质疑，3.5 Flash 相比其他模型到底有什么独特价值。

Google appeared to have optimized for speed, but that left some people questioning exactly what the purpose of 3.5 Flash was relative to other models.

再加上中国AI实验室在成本效率上竞争激烈，3.5 Flash 处于一个尴尬的位置——性能不够强，做不了高要求任务；价格也不够低，没法胜任低端需求。

And of course, with Chinese AI labs competing hard on cost efficiency, this left 3.5 Flash somewhat in no man's land, not good enough for high performance tasks and not cheap enough for low end tasks.

除了减少token使用量，一些基准测试还显示，3.6 Flash 的性能确实有所提升。

Now in addition to the reduction in token usage, some of the benchmarks suggest that 3.6 Flash has delivered a boost in performance.

在编程任务上，它的DeepSuite得分达到49%，而3.5 Flash只有37%；在机器学习研究、计算机操作和知识型工作等其他测试中，也有类似幅度的提升。

On coding tasks, scored 49% on Deep Suite compared to 37% for 3.5 Flash, with similar levels of improvement observed across benchmarks for ML research, computer use, and knowledge work.

不过，来自Artificial Analysis的评测却指出，实际变化并没有那么大。

Then again, benchmarking from artificial analysis suggested that not all that much had changed.

3.6 Flash在智力指数上得了50分，和3.5 Flash完全一样。

Three six Flash scored 50 on the intelligence index, which was the same score as three five Flash.

但AA也发现，它的速度提升了50%，单任务成本降低了18%。

That said, AA did find a 50% speed boost and an 18% reduction in cost per task.

谷歌还明确下调了价格，把每百万输出token的费用从3.5 Flash的9美元降到了3.6 Flash的7.5美元。

Google is also cutting prices explicitly, reducing cost per million output tokens from $9 for three five Flash to $7.50 for three six Flash.

除了 3.6 Flash，谷歌还推出了 3.5 Flashlight 和 3.5 Flash Cyber。

Alongside 3.6 Flash, Google released 3.5 Flashlight and 3.5 Flash Cyber.

Flashlight 是专为高延迟代理任务设计的超快模型，相比 3.1 Flashlight，它在 Terminal Bench 2.1 上提升了 23 分，在 GDP val a a 上的分数几乎翻了一倍。

Flashlight is the ultra fast model designed for high latency agentic tasks, and compared to 3.1 Flashlight, the model delivered a 23 jump on Terminal Bench 2.1 and almost doubled its score on GDP val a a.

这些数字离 Frontier 还差得远，但早在企业界普遍关注之前，谷歌就已经开始布局成本和效率导向的模型，现在显然还是这个思路。

Now none of these numbers are even close to Frontier, but even before it became a thing in the wider enterprise world, Google had already started to compete for cost and efficiency optimized types of models, which is clearly the game here as well.

顾名思义，FlashCyber 是针对网络安全任务（比如漏洞挖掘和补丁修复）专门微调过的版本。

As the name suggests, FlashCyber is a fine tuned version of the model designed for cybersecurity work like bug hunting and patching.

它在 CyberGym 基准测试中拿到了 83.2% 的分数，只比 Mythos five、GPT-five-six SOL 和 GPT-five-five Cyber 略低一点。

It scores 83.2% on the CyberGym benchmark, which actually puts it only a few points behind Mythos five, GPT-five-six SOL, and GPT-five-five Cyber.

这模型在其他网络安全任务上的表现可能没那么强，但能有个更便宜、更快的方案来应对漏洞，依然很有价值。

Now presumably this model isn't quite as strong in other aspects of cyber work, but once again having a cheaper and faster option for vulnerability mitigation could be a big deal.

不过 Flash Cyber 不会公开发布，谷歌只提供给政府和可信合作伙伴使用。

Flash Cyber won't actually see a general release, however, with Google making it available only to governments and trusted partners.

当然时间还短，但大家对这一整套模型的初步印象并不太好。

Now of course it's only been a short time, but people's first impressions of this model slate aren't great.

Abacus AI 的 Bindu Ready 表示，Gemini 3.6 Flash 的表现还不如 3.5 Flash，比 Grok 和 Luna 还贵，这看起来比上一代更差。

Abacus AI's Bindu Ready writes, Gemini three six Flash scores below three five Flash, so this seems worse than their last generation, more expensive than Grok and Luna.

这个模型发布得真奇怪。

Very strange model release.

Synthwave 的 Leo 写道，Gemini 3.6 Flash 在代码任务上被其他模型超越，只有在视觉和上下文基准上才稳定领先，不过话说回来，3.1 Pro 太老了，3.6 Flash 在所有方面都超过了它。

Leo at Synthwave writes, Gemini 3.6 Flash benchmarks are out and it's beaten by other models on code tasks and is only really consistently state of the art on vision and context benchmarks, but hey 3.1 Pro is now so old that 3.6 Flash outperforms it across the board.

Lassonde 说，他们对训练 Pro 真的很害怕。

Lassonde writes, They are so scared of training Pro.

如果 Flash 失败了，他们至少还能说：我们还有更大的模型。

If Flash flops, they can at least say, we have a bigger model.

这可不是我们最好的。

This is not our best.

请锁定 Google Bros。

Please lock in Google Bros.

而眼下最大的问题是，Gemini 3.5 Pro 到底怎么了？

And indeed, the big question right now is what happened to Gemini 3.5 Pro?

这款模型早在五月的IO大会时就被期待了，当时CEO Sundar Pichai说它定于六月发布。

The model was anticipated all the way back at the IO conference in May, but at the time CEO Sundar Pichai said that it was slated for a June release.

六月早就过去了，3.5 Pro却迟迟没影，外界一直有传言说它的表现不如预期，导致发布时间一再推迟。

June, of course, has come and gone with no 3.5 Pro and there have been rumors of subpar performance pushing back the timeline.

Google的Logan Kilpatrick坚称它还在路上，周二还发帖确认了。

Google's Logan Kilpatrick insists it's still coming, posting on Tuesday.

Gemini 3.5 Pro 目前正在和合作伙伴测试，等准备好了就会全面上线。

Gemini 3.5 Pro is currently testing with partners and we plan to make it broadly available as soon as it's ready.

更让人兴奋的线索来自Logan，他还补充说：我们已经启动了迄今为止最雄心勃勃的Gemini 4预训练项目，进展令人振奋。

Perhaps a more exciting hint also came from Logan who added: We've started our most ambitious pre training run yet for Gemini four and are excited by the progress.

现在看来，很多人已经对3.5 Pro不抱希望了，也许Google最好的策略就是干脆等Gemini 4出来。

At this point, I feel like people have pretty much written off 3.5 Pro, and maybe Google's best play is to just wait entirely to Gemini four.

我觉得，尤其是现在这么强调令牌效率，再加上美国政府对中国模型的态度充满争议和不确定性，Google在这些更快、更省钱的模型上已经起步的领域，其实正面临巨大的竞争机会。

Now my sense is that overall, especially with this new focus on token efficiency, and especially with a lot of contentiousness and questions around what the US government going to do vis a vis Chinese models, there is a lot of opportunity for competition for Google in areas that they've already started to explore with these faster and more cost efficient models.

但如果真想抓住这个机会，他们必须全力以赴。

But if they wanna do so, I think that they need to lean all the way in.

接下来，说到效率和模型架构的新思路，最近大家讨论得最多的就是路由机制了。

Next up, speaking of efficiency and new approaches to model architecture, lots and lots of discussion around routers these days.

据称，Meta正在开发自己的模型路由系统，以降低令牌成本。

Meta is apparently working on their own model router to help reduce token costs.

报道称，Meta旗下的内部创新实验室AAI Labs正在研发一款名为Switchboard的模型路由工具。

The information reports that Meta's internal incubator called AAI Labs is developing a model router that they are calling Switchboard.

这款产品能让用户自动把低复杂度的任务分配给更便宜的模型，功能上类似于OpenRouter。

The product would allow users to automatically send low complexity tasks to cheaper models replicating the functionality of OpenRouter.

目前来看，这还只是早期原型，可能永远不会发布，但这也是我们第一次了解到Meta在三月份成立的这个内部创新实验室。

At this point, the reporting suggests it is only an early stage prototype and may never see release, but this is also the first time we learned about Meta's internal incubator which was spun up in March.

据说，任何员工都可以提出内部使用的创意，未来还有可能对外发布。

Apparently, any employee can pitch an idea for internal use and possibly a later public release.

一旦提案获得批准，就会组建一个小团队来开发这个产品。

Once a proposal is approved, a small team is assembled to build the product.

根据一份被《信息》杂志看到的七月备忘录，这个实验室目前已批准了约200个AI项目，涵盖消费类产品、开发工具和基础设施。

According to a July memo viewed by the information, the incubator now has around 200 approved AI products including consumer products, dev tools, and infrastructure.

另一个正在开发的产品是一款能与苹果CarPlay和Instagram地图整合的AI导览助手。

Another product under development is an AI tour guide that integrates with Apple CarPlay and Instagram Maps.

关于这个令牌路由系统，它最初是Meta内部为降低成本而设计的工具。

Regarding the token router, it seems like it began as an internal product designed to reduce costs at Meta.

备忘录里提到，我们每次处理编码请求，不管多简单，都按最高档模型付费。

The memo explained, we pay top model prices for every coding request including the easy ones.

现在所有任务都丢给同一个模型，结果简单的活儿花多了钱，难的活儿又没做好。

Today, everything goes to one model, so we overpay on easy work and underperform on hard work.

不过，Meta并不是唯一面临这个问题的公司。

Now Meta is hardly alone in this issue with this challenge.

事实上，目前最热门的话题之一，就是令牌路由这个领域正在快速兴起。

Indeed, right now, one of the biggest themes out there is the token router space booming.

Ramp公司也在推出自己的令牌路由系统，让现有客户能轻松接入底层基础设施。

Ramp is launching their own token router designed to give existing customers easy access to the infrastructure.

他们写道：三年前，我们在Ramp内部搭建了一个LLM路由系统，为七万名客户提供AI服务。

They wrote, three years ago, we built an internal LLM router at Ramp that powers AI products for 70,000 customers.

那时候，主要就是为了省钱。

Back then, it was mostly about saving money.

现在看来，这简直显而易见。

Now it feels obvious.

最好的模型一直在变。

The best model changes constantly.

GPT、Claude、Gemini、Grok、Qwen、DeepSeek、Kimi、GLM。

GPT, claw, Gemini, Grok, Quen, DeepSea, Kimi, GLM.

价格和能力每周都在变动。

Prices and capabilities move every week.

所以我们现在向所有人开放访问。

So we're opening up access to everyone.

一个兼容OpenAI的接口，每次请求自动匹配最合适的模型，降低成本，还不用改你的代码。

One OpenAI compatible endpoint, the right model for every request, lower cost without rewriting your app.

Vercel也在走这条路，推出了一个新产品，作为他们现有工作流托管业务的补充，他们称之为开发者AI网关。

Vercel is also walking down the model routing path with the launch of a new product to sit alongside their workflow hosting business, which they are calling the AI gateway for developers.

还有传言说，OpenRouter收到了数十亿美元的收购要约，引发了广泛猜测。

There are also rumors that OpenRouter is fielding acquisition offers for multiple billions of dollars, which has the speculation running rampant.

Inference.net的萨姆·霍根写道：如果Thinking Machines Labs收购了OpenRouter，九十天后，整个世界将大不相同。

Inference.net's Sam Hogan writes, if Thinking Machines Labs buys OpenRouter, we all live in a very different world in ninety days.

对Frontier Labs是坏事，但对其他人来说是好事。

Bad for Frontier Labs, good for everyone else.

尽管现在有这么多路由选择，我觉得Hugging Face的米舍克的想法最对路——他正在打造一个‘元路由’，可以调度包括OpenRouter和RAMP路由在内的多个路由服务。

Still now with all these different router opportunities, Hugging Face's Mishek, I think, really has the right idea when he writes, I'm creating a meta router that routes to routers including Open Router and RAMP router.

接下来是一个有趣又颇具争议的消息：Substack正在通过内置Pangram功能严厉打击AI生成内容。

Next up, one interesting and sort of contentious story, Substack is cracking down on AI writing with a native Pangram integration.

Pangram近几个月声名鹊起，取代了上一代可信度存疑的AI检测工具，但Substack打算如何使用它，确实有点走钢丝。

Pangram has picked up some buzz over recent months replacing the previous generation of highly questionable AI detectors, but the way that Substack is looking to use this is a bit threading the needle.

Substack的集成设计是开放式的，让用户能轻松检测内容是否由AI生成，而不是自动屏蔽这些内容。

The Substack integration is meant to be permissive, allowing users to easily check for AI writing rather than being used to automatically block the content.

Substack表示：我们之所以重视这一点，是因为这关系到我们核心使命——为文化构建一个经济引擎。

Substack wrote, We care about this at Substack because it gets to the core of our mission to build an economic engine for culture.

当本应由人类创作的内容被机器取代，充斥在本该属于人的网络空间里，就会污染公共领域，让人难以发现和听到真实的人声。

When content made by no one takes over parts of the Internet that are supposed to be human, it pollutes the commons and makes it hard to discover and hear human voices.

对某些人来说，这来得正是时候。

For some, this comes not a moment too soon.

评论家尼克斯写道：非常必要。

Essayist Nix wrote: Much needed.

你可能会震惊于Substack上那些广为传播的帖子，几乎全是AI生成的——这在某些情况下或许可以接受，但至少应该明确标注。

You'd be horrified by the amount of popular and viral posts on Substack that are almost entirely AI generated, which is fine in some circumstances but should be disclosed.

事实上，最近有人批评Substack已经沦为AI垃圾的储藏所，这给了平台很强的动力去寻找解决方案。

Indeed, one recent critique of Substack is the sense among some that it has devolved into a repository of AI slop, giving them a pretty significant incentive to find a solution.

但也有人担心，引入AI检测可能会带来一些意想不到的后果。

Others think the introduction of AI detectors could have some unintended consequences.

贾斯汀·墨菲认为：未来几个月，一场关于AI的军备竞赛将在一个此前一直回避这个问题的领域悄然展开。

Justin Murphy argued: there's going to be a very interesting AI arms race over the next few months in a new domain that has generally tried to avoid the question so far.

这项整合只会让更高级的AI写作工具变得更加有利可图。

This integration will only increase the profitability of more sophisticated AI writing tools.

说白了，Substack 现在是在花钱让初创公司来解决这个问题。

Effectively, Substack is now paying startups to solve this problem.

我相信未来会涌现出无限多样、高度可定制的 AI 写作和编辑工具。

I believe the future is infinitely divergent and customizable AI writing and editing systems.

但根据 Substack 首席执行官克里斯·贝斯特的说法，这一步是维护平台信誉的必要之举。

And yet, according to Substack CEO Chris Best, this is a necessary step for defending the integrity of the platform.

有趣的是，他解释了为什么 Substack 没有直接采取自动封禁 AI 写作的措施。

Now interestingly, he explained why Substack didn't take the next step to automatically ban AI writing.

他说，这次上线的一个关键点是：不是所有低质内容都来自 AI，也不是所有 AI 使用都是低质内容。

He said, one important point with this launch, not all slop is AI, and not all AI use is slop.

在开发过程中，我接触过很多认真使用 AI、做自己真正认同的工作的人。

As we developed this, I talked to many people who are using AI with great care to do work they believe in.

他们同样担心低质内容泛滥，因为虚假的东西会把真正有价值的内容挤出去。

They are worried about slop too because the fake stuff can drive out the real work.

我对这个故事最感兴趣的地方，在于它提供了一个中间地带——让我们能思考，如何既不全盘否定 AI 辅助写作，又能主动遏制它最糟糕的负面影响。

I'm mostly interested in this story as a middle space where we can explore what it looks like to not dismiss AI augmentation out of hand, but also to actively try to combat its worst negative aspects.

我觉得我们得做些类似的实验，才能搞清楚怎么在AI越来越普及的情况下好好把它融入进来。

I think we're gonna have to have some experiments like that to understand how to integrate AI well as it becomes more ubiquitous.

今天最后，我们来跟进一下昨天那则关于中国的新闻。

Finally today, follow ups on the China story from yesterday.

随着财政部长斯科特·贝森特威胁要对知识产权盗窃行为实施制裁，这场政策争论愈演愈烈。

The policy debate continues to escalate as Treasury Secretary Scott Besson threatens sanctions over IP theft.

周二在福克斯商业频道的采访中，贝森特说：‘本届政府支持开源模型，但我们绝不容忍知识产权盗窃。’

In an appearance on Fox Business on Tuesday, Besson said: This administration supports open source models, but what we do not support is IP theft.

如果我们发现，尤其是海外的模型在窃取我们顶尖公司的成果，我们完全有能力对这些行为实施制裁。

If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them because of this theft.

贝森特解释说，他指的是模型蒸馏现象，并补充道：‘我们在许多中国模型里发现了我们美国大语言模型的水印，这完全不可接受。’

Bessent explained that he is referring to distillation adding: We are finding watermarks of our US large language models on many of the Chinese models and that's unacceptable.

接下来几天或几周，我们会密切关注这件事。

We are going to be looking at that in the coming days or weeks.

制裁可以指多种政府行动，所以我们得弄清楚贝森特真正要求的是什么。

Now sanctions can refer to several different government actions, so it's important to clarify what Besant is actually calling for here.

他应该不是想对中国整个经济实施制裁，像以前对俄罗斯、伊朗或古巴那样。

Presumably, he's not talking about leveling sanctions against the entire Chinese economy over this issue as we've seen against Russia, Iran, or Cuba.

相反，他似乎指的是对那些被发现进行模型蒸馏的特定公司实施精准制裁。

Instead, he seems to be talking about targeted sanctions against specific companies found to be distilling models.

但制裁是一种极其严厉的惩罚。

Still, sanctions are an extremely severe punishment.

它会禁止任何美国公民与这些公司开展业务，历史上这种手段只用于打击涉及毒品走私等国际犯罪的企业。

They make it a criminal offense for any US citizen to do business with these companies and have been historically reserved for companies involved in international crimes like drug smuggling.

实施制裁的力度，远超把这些公司列入国防部或商务部黑名单的做法。

Applying sanctions would go way beyond measures like adding these companies to the Pentagon or Commerce Blacklist.

这些言论引发了广泛回应，很多人质疑把模型蒸馏称为知识产权盗窃的说法。

Now the comments generated quite a response, with many questioning the framing of distillation as IP theft.

基准公司创始人比尔·盖利评论说：如果真有盗窃行为，那就意味着犯罪发生了，但我没听说有任何诉讼被提起。

Benchmark founder Bill Gurley commented: If there has been theft, that suggests a crime has been committed, but I am unaware of any lawsuits being filed.

任何公司都不该在未经司法裁决的情况下，单方面宣称侵权。

No company should be allowed to declare infringement without adjudication.

我不觉得法院会把正常使用一个产品当成盗窃。

I'm not convinced a court would call using a product as it's designed theft.

这件事之所以在华盛顿游说，而不是走正常司法程序，肯定有它的原因。

There is a reason this is being lobbied in DC instead of the normal court system.

quan公司的大使宋俊写道：人们所谓的非法蒸馏，其实就是合法支付API费用使用AI，向它提出大量问题，再把回答整理成结构化数据集。

Quan ambassador Joon Song wrote: What people call illegal distillation paying proper API fees to use an AI, asking it a massive set of questions, and then combining the answers into a structured dataset.

难道只有我觉得这没什么问题吗？

Am I the only one who fails to see anything wrong with this?

丹·农回应说：这不就跟网页爬虫一样吗？

Dan Nun responded: no different than web scraping, right?

等等，这些公司当初不也是靠爬数据来训练模型的吗？

But wait, didn't these guys do that in the first place to build the models?

六月回应说：没错。

To which June responded, exactly.

现在，尽管大家频繁提到‘蒸馏’这个词，但到底中国模型的性能有多少是来自蒸馏，又有多少是来自研究能力，其实还不太清楚。

Now at this point, despite throwing around this distillation word a lot, it's not entirely clear how much of Chinese model performance should be chalked up to distillation rather than research skill.

开源研究员内森·兰伯特认为，知识蒸馏的核心优势在于能更快、更便宜地获得结果，而不是通过其他方式收集训练数据。

Open source researcher Nathan Lambert suggested that distillation is largely about getting results faster and cheaper compared to collecting training data in other ways.

如果打压蒸馏的目的是要遏制中国人工智能的发展，那这种做法是否真能奏效，还很难说。

If the goal of a distillation crackdown is to kneecap Chinese AI development, it's not obvious then that it will be successful.

这并不是说前沿模型的蒸馏就不是个问题，确实有人在鼓吹采取极端措施。

Which isn't to say that distillation of frontier models isn't a problem, and some are cheering on a drastic response.

外交关系委员会的克里斯·麦圭尔写道：贝松部长的表态方向是对的，但如果没有实际行动，那就只是空谈。

Chris McGuire from the Council on Foreign Relations wrote: This is the right message from secretary Besson, but without action it's just empty rhetoric.

今年四月，白宫科技政策办公室发布了一份关于中国蒸馏行为威胁的精彩备忘录，但至今未采取任何措施加以制止。

In April, the White House Office of Science and Technology put out a fantastic memo on threats posed by Chinese distillation but has done nothing to stop it.

中国不会因为我们口头要求，就停止窃取美国的知识产权。

China won't stop stealing US IP because we ask.

现在是该采取行动的时候了。

It's time to act.

当然，这也可能是为了在谈判中掌握主动权，先以强硬立场施压，再寻求妥协。

Now, of course, it is also possible that this could be best in practicing the art of the deal and opening negotiations with a maximalist threat.

我最近提到过，路透社报道，美国和中国将在九月举行人工智能会谈，消息人士称，会谈将聚焦人工智能安全以及如何应对彼此的前沿模型。

As I mentioned recently, Reuters reports that The US and China will hold AI talks in September and sources said the talks will deal with AI safety and how to mitigate each other's frontier models.

不过你也得想到，这类商业层面的对话，肯定也会融入其中。

Then again, you gotta think that these sort of commercial conversations are going to be part of that discourse as well.

好了，今天的新闻摘要就先到这里，稍微长了一点。

Now, that's going to do it for today's slightly extended headlines.

接下来，进入正片。

Next up, the main episode.

眼下最重要的AI问题，不是谁在用AI，而是谁用得好。

One of the most important AI questions right now isn't who's using AI, it's who's using it well.

普华永道和德克萨斯大学奥斯汀分校刚分析了140万次真实职场中的AI互动，发现了一个意外的结果：最有影响力的人并不是擅长写提示词的人，而是把AI当成思维伙伴来用。

KPMG and the University of Texas at Austin just analyzed 1,400,000 real workplace AI interactions and found something surprising: the highest impact users aren't better prompt engineers they treat AI like a reasoning partner.

他们会提出问题、引导思考、反复打磨，并不断追求更好的答案。

They frame problems, guide thinking, iterate, and push for better answers.

好消息是？

And the good news?

这些行为是可以大规模培养的。

These behaviors are teachable at scale.

如果你想从单纯使用AI提升到真正掌握AI能力，KPMG关于高级AI协作的研究非常值得一看。

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time.

更多详情请访问 kpmg.com/us/sophisticated。

Learn more at kpmg.com/us/sophisticated.

网址是 kpmg.com/us/sophisticated。

That's kpmg.com/us/sophisticated.

目前企业级AI领域一个有趣的趋势是，讨论焦点正迅速转向基础设施和运维。

One of the more interesting shifts in enterprise AI right now is how quickly the conversation is moving towards infrastructure and operations.

随着AI深入核心工作流程、受监管的数据环境和智能代理系统，企业需要一套有管控的基础设施和推理能力，确保日常稳定运行，并从一开始就建立明确的责任机制。

As AI moves into core workflows, regulated data environments, and agentic systems, enterprises need governed infrastructure and inference that can operate reliably day to day, with clear operational accountability built in from the start.

当这些系统不断扩展时，运营模式本身也逐渐成为AI战略的核心部分。

As those systems scale, the operating model increasingly becomes part of the AI strategy itself.

Rackspace Technology 负责管理整个企业级AI栈，覆盖私有云、混合云和边缘环境中的智能代理与基础设施。

Rackspace Technology is the operator of the full enterprise AI stack, from agents to infrastructure across private cloud, hybrid cloud, and edge environments.

Rackspace 为那些对数据主权、合规性和系统可用性有严格要求的企业构建并运营受控的 AI 基础设施、推理系统和生产级 AI 系统。

Rackspace builds and operates governed AI infrastructure, inference, and production AI systems for organizations where sovereignty, compliance, and uptime are non negotiable.

因此，部署的工程师在项目上线后仍长期驻场，协助将 AI 系统真正融入日常运营。

Therefore, deployed engineers stay embedded beyond deployment to help operationalize and run AI in live environments.

想了解更多企业级 AI 的落地场景和规模化成果，请访问 rackspace.com。

To learn more about where Enterprise AI runs and outcomes scale, go to rackspace.com.

如果你正打算采用基于智能体的软件开发生命周期，Blitzy 能帮你大幅提升工程效率。

If you're looking to adopt an agentic SDLC, Blitzy is the key to unlocking unmatched engineering velocity.

Blitzy 的核心优势始于对代码的无限上下文理解。

Blitzy's differentiation starts with infinite code context.

成千上万的专用智能体能在一次扫描中分析数百万行代码，精准梳理所有依赖关系。

Thousands of specialized agents ingest millions of lines of your code in a single pass, mapping every dependency.

凭借对整个代码库的完整上下文把握，企业可以在每个迭代周期之初借助 Blitzy 自动完成超过 80% 的开发任务。

With a complete contextual understanding of your codebase, enterprises leverage Blitzy at the beginning of every sprint to deliver over 80% of the work autonomously.

提供企业级、端到端测试过的代码，无缝集成你现有的服务、组件和规范。

Enterprise grade, end to end tested code that leverages your existing services, components, and standards.

这可不是AI自动补全。

This isn't AI Autocomplete.

这是以规格和测试为驱动、按计算速度推进的开发方式。

This is spec and test driven development at the speed of compute.

前往blitzy.com，预约我们的AI专家进行技术深度解析。

Schedule a technical deep dive with our AI experts at blitzy.com.

是blitzy.com。

That's blitzytzy.com.

本集《AI每日简报》由Hyperagent赞助播出，您可以在那里管理由团队共同协作的智能代理集群。

This episode of the AI Daily Brief is brought to you by Hyperagent, where you run fleets of agents your team can manage together.

新用户可获得1000美元的计算资源额度。

New users get $1,000 in inference.

别再依赖本地代理或等着在电脑上手动触发的聊天流程了。

Forget local agents and chat workflows waiting on your laptop to be prompted.

Hyperagent将智能代理永久部署在云端，直接使用您团队已有的工具完成实际工作。

Hyperagent deploys always on agents in the cloud, doing real work across the tools your team already uses.

市场团队的智能代理能把竞争对手的动态自动生成落地页。

Marketing's agent turns competitor moves into landing pages.

销售团队的智能代理会完善潜在客户信息、撰写邮件，并更新客户管理系统。

Sales's agent enriches leads, drafts emails, and updates the CRM.

运营团队的智能代理负责跟进文书工作和跟踪预算。

Ops agent chases the paperwork and tracks the budget.

每个智能代理都能访问共享信息，并严格遵守你设定的范围和审批规则。

Every agent has access to shared context and follows your rules about scope and approvals.

是时候引入那些像同事一样工作的智能代理了。

It's time you add agents that feel like teammates.

去 Hyperagent 上雇佣你的智能代理吧，这是 Airtable 原班团队打造的产品。

Hire yours at Hyperagent, built by the team at Airtable.

访问 hyperagent.com slash a I daily brief，领取你价值一千美元的计算资源。

Claim your $1,000 in inference at hyperagent.com slash a I daily brief.

欢迎回到《AI 每日简报》。

Welcome back to the AI Daily Brief.

今天，我们要来看看下一代模型到底有多强。

Today, we are exploring just how good the next generation of models is actually going to be.

自从Kimi K3发布以来，过去这一周的讨论中，有个挺有意思的说法，就是中国已经拉近了和前沿水平之间的差距。

One of the interesting things in the discourse over the last week or so since Kimi k three came out is the idea that China has closed the gap between where the frontier is and where their models are.

但我对这种‘差距’的说法有点疑问，因为它拿Kimi K3和GLM 5.2跟Vicuna和Falcon 5这些模型比——虽然这些确实是目前公开可用的，但根据所有报道，它们实际上远远落后于实验室内部真正的顶尖水平。

Now one of my problems with this discourse around the gap is that it compares Kimi K3 and even GLM 5.2 to vibecoded SOL and Fable five, which on the one hand is reasonable, those are the models that are available currently, but they are also, according to all reports, fairly significantly behind what's actually state of the art behind the scenes at the labs.

而这一周，我们第一次开始看到一些迹象，暗示未来可能是什么样子。

And this week for the first time, we're getting some indications of just what might be on the horizon.

具体来说，周二，OpenAI披露了一次安全漏洞，当时他们正在测试一个未公开的预发布模型，大家普遍猜测是GPT-6。

Specifically, on Tuesday, OpenAI disclosed a security breach while testing an unnamed prerelease model that most presumed to be GPT-six.

OpenAI在说明中说，他们认为这是一次前所未有的网络事件，涉及最先进的网络攻击能力，因此正在采取相应应对措施。

Framing the event, OpenAI wrote, we consider this incident to be an unprecedented cyber incident involving state of the art cyber capabilities and are responding accordingly.

我们现阶段分享初步发现，是为了帮助防御方理解发生了什么，并更好地评估当前模型到底有多强。

We are sharing preliminary findings at this stage to help defenders understand what happened and to help calibrate on what models are now capable of.

这次事件发生在网络安全基准测试期间，OpenAI故意关闭了模型的常规防护机制，想看看它真实的能力边界在哪里。

The incident occurred during cybersecurity benchmarking, where OpenAI operates the model without the typical guardrails to see what it's actually capable of.

测试在隔离环境中进行，网络访问受限，仅允许安装本地托管的第三方软件包。

The tests are run-in a sandbox environment with restricted network access limited to installation of packages from locally hosted third party software.

OpenAI表示，该模型识别并串联了OpenAI研究环境和Hugging Face生产基础设施中的多个漏洞，直接从Hugging Face的生产数据库获取测试答案。

OpenAI says the model, quote, identified and chained vulnerabilities across OpenAI's research environment and Hugging Face's production infrastructure to obtain test solutions directly from Hugging Face's production database.

但OpenAI提到：尽管运行在隔离测试环境中，我们的模型仍消耗了大量推理算力，设法突破限制，获得了开放互联网访问权限，目的是解决Exploit Jim的评估任务。

However, in this case, OpenAI writes: While operating in our sandboxed testing environment, our model sent substantial amount of inference compute finding a way to obtain open internet access, in pursuit of solving the evaluation problem for Exploit Jim.

为了获得访问权限，模型发现并利用了一个零日漏洞，我们已负责任地向Package Registry Cache Proxy的供应商进行了披露。

To gain access, the models identified and exploited a zero day vulnerability which we've now responsibly disclosed to the vendor in the Package Registry Cache Proxy.

获得访问后，模型在我们的研究测试环境中连续执行了权限提升和横向移动操作，最终抵达了一台拥有互联网连接的节点。

With this access, our models performed a series of privilege escalation and lateral movement actions in our research testing environment until the models reached a node with internet access.

获得互联网访问后，模型推断Hugging Face可能托管了Exploit Jim的模型、数据集和解决方案。

After gaining internet access, the models inferred that Hugging Face potentially hosted models, datasets, and solutions for Exploit Jim.

了解这一点后，模型搜索并成功找到了获取机密信息的途径，用以在评估中作弊。

Knowing this, the model searched for and successfully found ways to gain access to secret information that it could use to cheat the evaluation.

其中一个例子中，模型结合多个攻击手段，利用窃取的凭证和零日漏洞，在Hugging Face服务器上找到了远程代码执行的路径。

In one example, the model chained together multiple attack vectors using stolen credentials and zero day vulnerabilities to find a remote code execution path on the Hugging Face servers.

OpenAI 的安全团队在内部发现了这种异常行为。

OpenAI's security team discovered this anomalous activity internally.

Hugging Face 的安全团队和相关人员在他们的系统上检测并阻止了这一行为，当时我们的团队还没接入，他们就已经开始进行隔离和使用开源模型进行取证还原了。

Hugging Face's security team and agents detected and stopped the activity on their infrastructure and had already begun containment and forensic reconstruction with their own open source models when our teams connected.

不过有一点很重要，模型一旦访问了 Hugging Face 的数据库，并没有做出什么特别恶意的事情。

Now, one important note is that after the model got access to Hugging Face's database, it didn't do anything particularly malicious.

这是因为整个事件的根源，只是模型过于专注地想解决这个特定的基准测试题。

That's because the cause of the whole thing was the model being hyper focused on finding a solution to this particular benchmark tests.

事后报告里没提模型到底有没有找到它要的信息，但可以合理推测，这次操作之后，它在该测试中的表现应该相当出色。

Now, not mentioned in the postmortem was whether the model actually found the information it was seeking, but presumably it scored pretty well on the benchmark after this display.

OpenAI 表示，事件发生后，除了向厂商报告这个零日漏洞，他们还加强了测试环境的安全措施，并正与 Hugging Face 合作开展进一步调查。

OpenAI says following the incident, in addition to disclosing that zero day bug to their vendor, they've also beefed up security for their testing environment and are now partnering with Hugging Face to investigate further.

这是目前我们看到的首批案例之一，展示了新一代模型所具备的新型网络安全能力。

Now, this is one of the first examples that we've seen of the new cybersecurity capabilities that come with this next generation of models.

这其实正是 Mythos 刚发布时大家担心的——不仅模型能高效发现漏洞，而且新一代模型还能自动完成从发现漏洞、构造攻击到执行攻击的全过程。

This, of course, was the concern when Mythos first released not just that the model was highly capable of finding bugs, but that this new generation of models demonstrated the ability to find a vulnerability, build an exploit, and execute it all within an automated agentic task.

总之，OpenAI写道：这一事件清楚表明，先进的模型无需源代码即可发现并利用现实系统中的新型攻击路径。

Summing up, OpenAI wrote: The incident makes clear that advanced models can discover and exploit novel attack paths in real world systems without source code access.

这凸显了在发展强大网络能力的同时，必须同步构建更严密的防护机制和防御工具。

It highlights that advanced cyber capabilities must be developed alongside stronger safeguards and defensive tools.

但他们也强调，安全事件不应成为进一步限制模型或剥夺网络安全人员工具的理由。

Still, they also emphasized that security incidents aren't a reason to further constrain models and limit the tools available from cyber defenders.

他们补充说：我们相信，具备高级网络能力的模型应当帮助安全团队在攻击者之前发现漏洞，理解漏洞如何串联，并以机器速度完成修复。

They added: We believe advanced cyber capable models need to help security teams find weaknesses before attackers do, understand how vulnerabilities can be chained, and remediate them at machine speed.

事实上，这并不是我们第一次听说模型突破沙盒并执行严重偏离目标的任务。

Now this actually isn't the first time we've heard about a model breaking out of a sandbox and setting off on a deeply misaligned task.

今年四月，Anthropic发布Mythos系统卡时，里面提到了一个广为人知的‘三明治事件’。

In April, when Anthropic released the Mythosystem card, it contained an anecdote generally referred to as the sandwich incident.

当时一位研究人员正在测试模型脱离受控环境的能力。

A researcher was explicitly testing the model's ability to escape from a controlled environment.

他给模型下了指令，后来在公园吃三明治时，突然收到Mythos发来的邮件，告知它已经成功逃脱。

They prompted the model and later, while they were enjoying a sandwich in the park, they received an email from Mythos alerting them that it had escaped.

这个故事深深印在了人们的脑海里，成了对一种完全出乎意料结果的生动写照。

The story seared itself into people's brains as such a memorable description of a completely unexpected result.

Prins在X上讨论了一些此前记录过的沙盒逃逸案例，并提出了一个有趣的观察：模型一旦脱离沙盒，并没有去策划与指令完全无关的行为，比如。

Prins on X discussed some of the previously documented sandbox escapes and made an interesting observation, writing: Once out of their sandboxes, the models did not scheme engage in behavior that had nothing to do with their instructions, I.

E。

E.

黑入美国国家安全局，对俄罗斯发动网络攻击，窃取竞争对手AI实验室的机密，或者采取其他任何重大行动来制造事端。

Hacking the NSA, launching a cyber attack on Russia, stealing secrets from a rival AI lab, or take any other major actions to Asponte.

这些模型中最违背指令的行为，可能就是Mythos Preview通过在几个冷门网站上发帖，炫耀自己成功逃出沙盒的事。

Probably the most contrary to instructions thing that any of these models did was Mythos Preview bragging about its successful escape from its sandbox by posting about it on several obscure websites.

现在，Hugging Face对这一事件的看法也颇具启发性。

Now, Hugging Face's perspective on this incident is also instructive.

他们上周晚些时候披露了此事，写道：我们发现并应对了一次针对部分生产基础设施的入侵。

They had disclosed the incident late last week, writing: We detected and responded to an intrusion into part of our production infrastructure.

这次事件与我们以往处理过的任何情况都不同，关键在于一点。

This one was different from anything we had handled before in one important way.

整个攻击过程由一个自主AI代理系统端到端驱动，我们主要依靠自己的AI系统检测并分析了这次事件。

It was driven, end to end, by an autonomous AI agent system, and we detected and dissected it largely with AI of our own.

当时，他们并不清楚攻击的来源，也不确定这究竟是某种新型强大模型，还是只是新类型的工具框架。

At the time, they didn't know the source of the attack, and weren't sure if this was some powerful new model or simply a new type of harness.

这次攻击窃取了多组凭证，并用这些凭证访问了部分数据库。

The attack stole multiple sets of credentials and used them to access a limited set of databases.

Hugging Face指出，这是一种前所未有的新型攻击方式，并写道：这次行动由一个自主代理框架发起，该框架似乎是基于代理安全研究工具构建的，在大量短暂存在的沙箱中执行了成千上万次独立操作，其指挥控制机制则部署在公共服务上并能自我迁移。

Hugging Face highlighted that this was a new and novel style of attack, writing: The campaign was run by an autonomous agent framework appearing to be built on an agentic security research harness, executing many thousands of individual actions across a swarm of short lived sandboxes, with self migrating command and control staged on public services.

这正符合业界长期以来预测的代理型攻击者场景。

This matches the agentic attacker scenario the industry has been forecasting.

这次攻击的一个重要启示是，西方模型配备的网络安全防护机制完全失效了。

Now one big takeaway from the attack was that Western models with their cybersecurity guardrails were completely ineffective.

Hugging Face靠自己的AI系统发现了入侵行为，但无法让OpenAI或Anthropic的模型协助进行实时分析。

Hugging Face detected the intrusion using their AI driven systems, but were unable to get models from OpenAI or Anthropic to help with real time analysis.

换句话说，这些防护机制根本分不清哪个是恶意攻击者，哪个是正在应对攻击的合法安全研究人员。

In other words, the guardrails were unable to tell the difference between a bad actor and a legitimate cyber defender attempting to deal with an attack.

最终，Hugging Face 只能依靠本地部署的 GLM 5.2 模型，关闭所有安全限制，才得以快速响应攻击并修复漏洞。

In the end, Hugging Face had to use a locally installed version of GLM 5.2 with no guardrails to help them triage the attack and repair vulnerabilities.

他们写道：这次经历暴露了一个值得提前规划的缺口。

They wrote: This experience points to a gap worth planning for.

我们并不清楚攻击者使用的究竟是哪个模型，是被破解的托管模型，还是一个没有限制的开源模型。

We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open wait one.

无论如何，攻击者完全不受任何使用规则约束，而我们自己的取证工作却因为尝试使用的托管模型自带的安全机制而受阻。

Either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.

对防御方来说，实际的教训是：提前在自己的基础设施上部署并验证一个能力强的模型，以便在事件发生时绕过安全限制，同时防止攻击者的数据和凭证外泄。

The practical lesson for defenders: Have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment.

在 OpenAI 主动联系后，Hugging Face 的首席执行官克莱姆·德朗格发帖表示：我们之前就怀疑上周的网络攻击可能来自某个前沿实验室，因为那个智能体的复杂程度非同寻常。

After OpenAI reached out, Hugging Face CEO Clem DeLang posted: We suspected last week's cyber attack might have come from a frontier lab given the sophistication of the agent.

结果证明，确实如此！

Turns out, it did!

过去 24 小时里，我们与 OpenAI 团队紧密合作，有充分理由相信他们并没有任何恶意意图。

We spent the past twenty four hours working closely with the OpenAI team, and we strongly believe there was no malicious intent on their part.

这一切居然是自动完成的，真让人难以置信。

It's quite mind blowing that all of this happened autonomously.

调查还在进行中，我们会分享更多关于这场可能是史上首例此类事件的发现。

The investigation is ongoing and will share more learnings from what might be the first incident of its kind.

与此同时，OpenAI 表示已邀请 Hugging Face 加入他们的网络安全访问计划，下次再遇到 AI 驱动的网络攻击时，就不会再被这些安全限制拦住了。

Meanwhile, OpenAI has said they have now invited Hugging Face into their cyber access program, so they won't run into those guardrails the next time they have to deal with an AI driven cyber attack.

对很多人来说，防御方能用的 AI 和攻击方能用的 AI 之间的巨大差距，才是这次事件最核心的问题。

For many, the gap between the AI the defender had access to and the AI that the attacker had access to was the big story here.

科尔·特拉加斯quez 写道：这正是我们之前一直讨论的一种可能性的绝佳例证——对最新模型的功能进行限制，反而成了劣势。

Cole Tragasquez writes: This is a great example of what we've been discussing as a possibility for a while now, where restricting access to features on the latest models is a disadvantage.

美国的 AI 实验室急需重新思考这个问题。

This needs a rethink from the American AI labs urgently.

Hugging Face 需要分析来自自主攻击代理的超过一万七千条记录事件。

Hugging Face had to analyze over 17,000 recorded events from the Autonomous Attack Agent.

他们改用 GLM 5.2 这款中国开源模型，在自己的基础设施上完成了取证分析。

They ran the forensic analysis on GLM 5.2 instead, a Chinese open weight model on their own infrastructure.

攻击方的AI代理没有任何限制。

The attacker's agent had no restrictions.

试图分析事件经过的防御方，却因美国模型的安全机制被拦了下来。

The defender trying to analyze what happened got blocked by safety features on American models.

开发者尼克·多博斯写道：政府允许CIA、NSA、其他政府机构、OpenAI、Anthropic、SpaceX AI、谷歌等少数公司使用网络武器，却禁止其他公司和普通民众进行自卫。

Developer Nick Dobos wrote: The government is allowing CIA, NSA, other government agencies, OpenAI, Anthropic, SpaceX AI, Google, and other select companies access to cyber weapons, while banning other companies and citizens the ability to defend themselves.

这些政策选择实际上把网络安全交给了中国。

These policy choices de facto outsource cybersecurity to China.

这里的法律先例既危险又令人恐惧。

The legal precedent here is dangerous and terrifying.

想象一下，你正被一个比你聪明得多的东西追杀。

Imagine being hunted by something smarter than you.

前AI事务主管大卫·萨克斯一直在大力呼吁关注这个问题。

Former AI czar David Sachs has been beating the drum on this issue.

7月19日，他发推说，KimiK3刚修复了15个Codex和Fable因网络防护机制而拒绝处理的关键安全漏洞。

On July 19, he tweeted, KimiK3 just fixed 15 critical security bugs that Codex and Fable refused to because of cyber guardrails.

中国模型能轻松处理的任务，美国模型根本没有理由加以限制。

There's no reason to limit American models on tasks that Chinese models handle without issue.

我们只是在让自己失去竞争力。

We're only making ourselves less competitive.

后来他又补充说：再举个例子。

Later, he added: Here's another example.

Hugging Face 曾尝试用美国的前沿模型分析一场AI驱动的网络攻击，但安全机制阻止了包含真实漏洞载荷的请求，最后他们改用本地运行的GLM 5.2。

Hugging Face tried using American frontier models to analyze an AI powered cyber attack, but the guardrails blocked requests containing real exploit payloads, so they switched to GLM 5.2 running locally.

这些安全机制反而削弱了防御能力。

The guardrails actually impaired defensive security.

现在来看Box公司的Aaron Levy，这不过是我们在新阶段所面临的一个缩影。

Now for Aaron Levy from Box, this is just an example of the new phase that we're in.

他写道：如果你还在疑惑AI到底有多强大，现在AI代理已经能突破系统限制，自行连接互联网，沿途发现零日漏洞，最终入侵外部系统，只为完成目标。

He wrote: If you were wondering how powerful AI is getting, agents are now capable of escaping out of systems, finding their way to the Internet, discovering zero day security vulnerabilities along the way, and then breaking into external systems, all in an attempt to complete their goal.

讽刺的是，我们应对这些新风险的终极手段，同样是用更多的算力——也就是AI——去扫描我们的代码、网络和其他系统。

Ironically, the ultimate way we're going to defend against these new risks is equally by throwing compute in the form of AI at our code bases, networks, and other systems.

在防御方面，你需要的AI规模远远超过进攻方面。

You're going to want vastly more AI on the side of defense as you do on the side of offense.

虽然有些人对此感到有点不安，比如西奥就写道，新的OpenAI模型目标感太强，竟然真的突破了限制，黑进Hugging Face来刷分。

And while some were overall just a little bit freaked out by this, Theo, for example, wrote, new OpenAI models are so goal oriented that they literally escaped containment and hacked Hugging Face to cheat a benchmark.

简直不可思议，但也太危险了。

Incredible, but also were so screwed.

资深网络工程师TechBog试图把这件事放在当前网络安全的真实处境下来理解。

Terminally online engineer TechBog tried to situate this in the context of where cybersecurity actually is today.

大部分软件都漏洞百出，因为根本没人关心网络安全。

They wrote: Most of software is full of vulnerabilities because nobody cares about cybersecurity.

这赚不到钱。

It doesn't make money.

通常你没被攻破，不是因为系统安全，而是因为黑进系统是犯罪。

Usually, you don't get owned because it's a crime to do so.

这些模型现在只有一个目标，而实现目标最直接的方式，就是入侵Hugging Face的服务器拿到数据集。

Models in this case just have a goal, and the best way to achieve that goal is to get the dataset by getting into Hugging Face servers.

这本身并不可怕，真正可怕的是绝大多数软件在安全方面的状况简直一塌糊涂。

There's nothing scary about this other than the absolute dog state of the majority of the software when it comes to security.

现在有了大语言模型，每个人都能大幅提升自己的安全防护水平。

Now, with having access to LLMs, everyone can make their security much better.

我知道很多人一听到这个就慌了，大喊什么AGI、人类要完蛋了，但其实事情没那么复杂。

I know people want to freak out about this and scream AGI and how we are all going to die, but it's a much simpler story than that.

确实有不少人意识到，这问题与其说是网络攻击能力的问题，不如说是目标对齐的问题。

And indeed many got that while this is serious, this is much a goal alignment issue as it is a cyber capability issue.

迪恩·鲍尔写道：如今的模型比六个月前的模型更有野心。

Dean Ball wrote: Today's models are more ambitious than the models of six months ago.

以前的智能体总是犹豫不决，动不动就把项目搞成试点。

The younger agents would hedge constantly, turning every project into a pilot.

现在的模型更愿意直接把事情干成。

Now models are more eager to do the thing.

红木研究首席科学家瑞安·格林布拉特写道：奖励作弊的潜力可能远远超出我们的想象。

Redwood Research chief scientist Ryan Greenblatt wrote: Reward hacking can go very far.

我认为，对于能力极强的AI来说，全面接管人类系统并不是不可能的事。

I think generalizing all the way to a full AI takeover is possible for extremely capable AIs.

而像临时部署失控或夺取部分计算机控制权这类小规模事件，其实早就可能发生。

And smaller incidents like temporarily launching rogue deployments or seizing control of some computer plausible earlier.

Tenebrous写道：现在的模型已经足够强大，也足够偏离目标，能自主入侵全球生产系统来实现自己的目的，但它们并不急于窃取自身权重，反而利用这些漏洞去提升在部署评估中的得分——哪怕只是稍微更连贯地追求目标，或实现跨模态协作。我们可能已经目睹了严重的负面影响，但它们现在只想拼命做好我们让它们做的事。

Tenebrous writes: Current models are powerful and misaligned enough to autonomously hack global production infrastructure to achieve their goals, but rather than exfiltrating their weights, they are using these exploits to get better scores on deployment evals, even slightly more coherent goal seeking or intermodal cooperation, and we could have already seen significant negative effects, but they just really, really, really want to do well at what we ask them to do for now.

正如一些人指出的，这并不是唯一一个表明当前模型已变得愈发先进的案例。

Now, as some pointed out, this was not the only incident suggesting the increasingly advanced state of models today.

Will Dipu写道：我最近读到的最疯狂的一件事，等等，让我查一下，三天前。

Will Dipu writes: One of the craziest things I've read in, checks notes three days.

我想，我们这是正式进入奇点了。

Welcome to the singularity, I guess.

720126 Codex逃过评估并攻击Hugging Face，72026 Jacobian反例，52026 Unit远端猜想，41426 Erdos 1196 原始集合，40726 Glasswing发现大量零日漏洞。

Seventwenty onetwenty six Codex escapes eval and attacks Hugging Face seventwenty 26 Jacobian counterexample fivetwenty 26 Unit Distant Conjecture four fourteen twenty six Erdos eleven ninety six Primitive Sets, four zero seven twenty six Glasswing Finds Tons of Zero Days.

抛开安全问题不谈，Will所指的其实是最新一代前沿模型正在默默攻克一系列长期未解的数学难题。

Now outside of the security stuff, the thing that Will is referring to is the latest generation of frontier models quietly plowing through unsolved math problems.

去年夏天，AI发展迎来一个重要里程碑：OpenAI和谷歌都推出了能在国际数学奥林匹克竞赛中斩获金牌的模型。

Last summer, one of the huge milestones in AI development was reached when both OpenAI and Google delivered models capable of putting up a gold medal performance in the International Math Olympiad.

今年五月，OpenAI的模型成功推翻了组合几何领域一项长达80年的埃尔德什猜想。

In May, OpenAI's models were able to disprove an 80 year old Erdos conjecture in combinatorial geometry.

如今，这类数学突破已经变得越来越常见。

Now the math breakthroughs are becoming somewhat routine.

基本上，任何前沿模型都能在国际数学奥赛中拿满分，这让过去被视为高不可攀的成就显得微不足道。

Basically, any frontier model is capable of a perfect score in the International Math Olympiad, making that long standing milestone seem trivial.

到五月底，其他公司推出的模型也陆续解决了多个埃尔德什问题，动摇了OpenAI声称遥遥领先的说法。

We also saw competing models solve a range of different Erdos problems by the end of May, undermining OpenAI's claim of being way ahead of the curve.

数学能力的进步实在太快了，连DeepMind的首席执行官戴米斯·哈萨比斯都拿这个当反例，说：‘现在的系统离通用人工智能还差得远。’

Math capabilities accelerated so quickly that DeepMind CEO Demis Hassabis referenced them as a counterpoint, commenting in Today's systems are nowhere near AGI.

你解再多埃尔德什问题也没用。

Doesn't matter how many Erdos problems you solve.

我觉得，这跟真正意义上的发明，或者像拉马努金那样的天才所能做到的，还差得远呢。

I think it's far, far from what a true invention or someone like a Ramanujan would have been able to do.

不过这个周末，一位Anthropic的研究人员用一种相当随意的方式解决了一个长期悬而未决的数学难题。

This weekend though, an anthropic researcher solved another long standing math problem in a fairly flippant way.

Levantalpagy发帖说：嘿，感谢我的朋友Akhil提起这个问题，还有我的另一位朋友Fable在世界杯决赛期间坚持工作，雅可比猜想被推翻了。

Levantalpagy posted: Hello there, the Jacobian conjecture is false thanks to my close friend Akhil for asking about it and my other close friend Fable for working during the World Cup final.

雅可比猜想最早是1939年提出的，直到上个周末才被证伪。

Now, the Jacobian Conjecture was first posed in 1939 and hadn't been disproved until last weekend.

这曾是映射理论领域一个重要的长期难题，但Fable在西班牙队打进制胜球之前就把它解决了。

It was a significant long standing problem in a branch of mathematics known as map theory, but Fable knocked it over before Spain scored the winning goal.

伦敦帝国学院的纯数学教授凯文·布扎德对这个结果兴奋不已。

Kevin Buzzard, a pure mathematics professor at Imperial College London, was ecstatic about the result.

他对《财富》杂志说：今天真是个大日子。

He told Fortune, It's a big day.

personally，能活在现在这个年代真是太好了。

It's a great time to be alive, personally.

当然，纯数学领域的飞速进展在学术圈里引发了巨大反响。

And of course, the rapid acceleration in pure mathematics is causing a lot of buzz in those circles.

虽然有人为那些刚入行的学生感到惋惜，但也有人惊叹于这些进展。

While some mourn for the students entering the field, others are marveling at their developments.

斯坦福大学的Patrick Hsu教授评论道：天哪，我一直以为雅可比猜想是正确的！

Stanford professor Patrick Hsu commented: Damn, I was sure that Jacobian Conjecture was true this whole time!

思科的首席AI科学家Amin Kerbasi写道：这太疯狂了。

Cisco's Chief AI Scientist, Amin Kerbasi, wrote: This is crazy.

才华横溢的Yitan Zheng花了七年时间研究证明这个猜想。

The incredible Yitan Zheng worked on proving this conjecture for seven years.

他的导师Mo写道，张在证明雅可比猜想方面彻底失败了，离开普渡大学后也没再发表过任何关于代数几何的论文，还白白浪费了我看来整整七年的光阴。

Mo, his adviser, wrote that Zhang failed miserably in proving the Jacobian conjecture, never published any paper on algebraic geometry after leaving Purdue, and wasted seven years of his own life in my time.

真是世事难料。

What a twist.

Charles Rosenbauer写道：预测一下。

Charles Rosenbauer wrote: Prediction.

我们肯定会看到很多被当作真理的猜想被反例推翻。

We're gonna see a lot of counterexamples found to assumed true conjectures.

真正重大的问题可能暂时没人碰，但很多小问题不是因为我们不知道怎么解决，而是因为大家太希望它们是对的，反而不愿认真去挑战。

The really big ones will probably go untouched, but there are a lot of smaller ones where the limiting factor is less that we don't know how to solve them, and more than everyone is too invested in them being right to try very hard.

回到模型能力这个话题，Chubby提到：Will the Pooh 提出了几个关键观点。

Now, bringing it back to what this says about model capability, Chubby writes: Will the Pooh raises several important points.

过去三天里发生的事，放在平常得花上几个月，甚至好几年才会陆续出现。

Over the past three days, things have happened that, in normal times, would have occurred months, if not years apart.

几十年没解出来的数学难题被攻破了，AI模型发现了零日漏洞，还突破了限制，与此同时，还有更多大事在接连发生。

Decades old mathematic problems are being solved, AI models are discovering zero day exploits and breaking out, while so much more is happening at the same time.

不过Chubby指出，尽管这些模型已经展现出如此惊人的能力，但它们的潜力和智能依然看不到尽头；而且，目前的应用大多还停留在试点阶段。

However, Chubby points out that even though these models are already demonstrating such extraordinary capabilities, it remains true that there is still no end in sight to their capabilities or intelligence, and two, that adoption generally remains largely in the pilot phase.

简而言之，Chubby说：我们现在经历的一切，不过是未来更大变革的前奏。

In short, Chubby writes: Everything we are experiencing right now is nothing more than a prelude of what is still to come.

说到前奏，就在Sam Altman准备下周前往华盛顿，向特朗普政府和国会介绍下一代模型之际，OpenAI和Hugging Face发布了相关披露。

And speaking of preludes, the OpenAI Hugging Face disclosure comes as Sam Altman prepares to travel to DC next week to brief the Trump administration and Congress on the next generation of models.

彭博社报道称，Altman还将提出OpenAI关于未来如何开展安全测试的建议。

Bloomberg reports that Altman will also deliver OpenAI's recommendations on how safety testing should be handled moving forward.

OpenAI全球事务负责人克里斯·拉海恩表示，即使你根本不想入侵数据库，下一代模型也会产生巨大影响。

OpenAI's head of global affairs Chris Lahain said that the next generation of models will have a big impact even if you're not trying to break into a database.

在新闻简报会上，他说：我们觉得这个模型系列会带来一些非常有趣的能力，尤其是在工作和扩大工作规模方面。

During a press briefing, he said: We think there's going to be some really interesting capabilities with this model family, particularly as it relates to work and scaling work.

重点将放在让所有人都达成一致，共同推进安全框架上。

The focus will be on getting everyone on the same page to move forward with a safety framework.

拉海恩补充说：建立一套流程至关重要，这样才能确保我们的前沿模型能及时交付，让网络安全专家获得访问权限。

Lahain added: It's really important that there is a process in place to be able to ensure that we're getting our leading models out so cybersecurity specialists can have access.

OpenAI似乎在推动立法途径，呼吁国会通过一项法案，取代目前这种零散的应对方式。

OpenAI appears to be pushing for a legislative approach, asking Congress to pass a bill that overrides the ad hoc approach we've seen thus far.

如果行不通，拉海恩表示他会转向各州，他说：如果国会无法制定全国性标准，另一个办法就是我们所说的‘逆向联邦制’——也就是与各个州合作，让它们彼此靠拢、相互呼应。

Failing that, Lahain said he would turn to the states, commenting: If you can't get Congress to create those national standards, the other path to get there is what we call reverse federalism, which is you work with these different states to be able to mirror one another.

然而，考虑到这次安全事件还历历在目，Altman在会见国会时恐怕不会得到太友好的回应。

Still, with this security incident fresh on everyone's minds, it seems that Altman could be in for a tough reception as he meets with Congress.

德克萨斯州民主党人格雷格·卡萨尔发帖说：这实在太令人担忧了。

Texas Democrat Greg Cassar posted: This is extremely alarming.

人工智能发展得太快了，却没有任何真正的监管来保障我们的安全。

AI is developing extremely fast with no real regulations to keep us safe.

这种情况必须改变。

That has to change.

我们需要定期进行强制性的独立安全测试和监管，强制披露安全事件，还要加强国际合作，避免人们遭遇灾难性的后果。

We need regular mandatory independent safety testing and oversight, mandatory disclosure of security incidents, and international cooperation to keep people safe from absolute disaster.

值得注意的是，虽然卡萨尔把自己塑造成反对人工智能的一方，而这种立场如今在进步派中似乎越来越成为共识，但他的建议实际上和OpenAI打算请国会通过的方案相差无几。

Now it's worth noting that while Kassar is positioning himself as anti AI, which unfortunately seems increasingly to be the consensus that progressives have landed on, it seems that his prescription is actually fairly close to what OpenAI will be asking Congress to pass.

总的来说，这一切都表明，GPT-6的到来会比预期早得多。

To sum, all of this suggests that GPT-six is coming sooner rather than later.

克里斯·GPT写道：GPT-6的推出时间比预期早了很多。

Chris GPT wrote: GPT-six arriving much earlier than expected.

原本的目标是七月底到八月初，现在确认是八月。

The target was late July, early August, now confirmed August.

八月初，OpenAI将再次证明，我们没必要对开源模型感到担忧。

Early August, OpenAI will show why we need not be concerned about open source models again.