Where Claude Opus 5 Fits in Your Model Rotation — 文稿

M10:00

今天在 AI Daily Brief 里，Anthropic 发布了 Claude Opus 5，我们会来聊聊它应该怎么放进你的模型配置里。

Today on the AI Daily Brief, Anthropic has released Claude Opus 5 and we are talking about where it should fit into your model setup.

在那之前，先看头条：关于 OpenAI 这个月早些时候对 Hugging Face 发起的失控模型攻击，外界还在持续追问。

Before that on the headlines, continued questions around OpenAI's rogue model attack of Hugging Face earlier this month.

AI Daily Brief 是一档每日更新的播客和视频节目，专门聊 AI 领域最重要的新闻和讨论。

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

好，朋友们，在我们正式开始之前，先快速说几个通知。

Alright friends, quick announcements before we dive in.

M11:07

好，现在来说上周的一条大新闻：OpenAI 对一个未具名模型做安全测试，而大家都猜那可能是 GPT-6。

Now, one of the big stories from last week revolved around OpenAI's security testing of an unnamed model, which people presumed to be GPT-6.

Hugging Face 和 OpenAI 都发布了这次攻击的事后复盘，从各自的视角讲了事情经过。

Both Hugging Face and OpenAI released postmortems on the attack telling the story from their view.

OpenAI 周三发布的博客文章表示，他们正在和 Hugging Face 密切合作，进行一次全面调查，这让人感觉两家公司关系还不错。

OpenAI's blog post released on Wednesday suggested that they were working closely with Hugging Face on a full investigation, implying the two companies were on good terms.

而就在今晚，Hugging Face 的 CEO Clément Delangue 正飞往 San Francisco，准备——用他自己的话说——和那个失控 agent 来一场小聊聊。

Tonight, Hugging Face CEO Clément Delangue was on a flight to San Francisco to have, as he put it, a little chat with that rogue agent.

在周六发的一篇后续帖子里，他写道：本着透明的精神，下面是我向 OpenAI 提出的要求。

In a follow-up post on Saturday, he wrote, In the spirit of transparency, here's what I asked OpenAI.

第一，彻底透明。

One, radical transparency.

把这个所谓“失控 agent”的 traces 公布出来，让整个研究社区都能研究到底发生了什么。

Let's release the traces from the quote-unquote rogue agent so the entire research community can study what happened.

第二，给防守方更多能力。

Two, more capability for defenders.

让 OpenAI 承诺拿出一亿算力，帮助 Hugging Face 社区用最好的开源和闭源模型，构建强大的网络防御能力。

Let's commit 100 million in compute from OpenAI to help the Hugging Face community build powerful cyber defenses with the best open and closed models.

这是第一次 autonomous agent 发起的网络攻击，是前所未有的事件。

The first autonomous agent cyber attack is an unprecedented event.

它值得一次前所未有的回应。

It deserves an unprecedented response.

而在 OpenAI 披露这起事件之后的这几天里，又出现了不少新闻报道，让整件事变得更加扑朔迷离。

Now in the few days since OpenAI disclosed the incident, we've had a number of news articles that add more confusion to the story.

The Wall Street Journal 写道，Hugging Face 对这次攻击完全措手不及，因为它看起来像是超人类级别的，超出了任何已知模型的能力范围。

The Wall Street Journal wrote that Hugging Face was caught completely off guard by the attack, which seemed to be superhuman and beyond the capabilities of any known models.

具体来说，这次攻击使用了一个复杂的 agent swarm 来躲避防御，在网络中移动时会迅速启动和关闭会话。

Specifically, the attack used a sophisticated agent swarm to evade defense, rapidly spinning up and shutting down sessions as it moved across the network.

一个很有意思的细节是，这场攻击整整持续了两天，Hugging Face 最后是在 GLM 5.2 的帮助下才把它关停。

One interesting detail was that the attack was ongoing for two whole days before Hugging Face was able to shut it down with the help of GLM 5.2.

现在，所谓“失控”这个说法，也就是模型的行为超出了 OpenAI 的控制，对这些媒体来说显然是最关键的概念。

Now this idea of rogue, that the model was acting beyond OpenAI's control, is definitely for these media outlets the key concept.

周五，Reuters 发了一篇报道，标题是：它的 AI agent 连续几天都在黑一家公司，但消息人士称 OpenAI 一周后才发现。

On Friday, Reuters dropped a piece titled, Its AI agent spent days hacking a company, but sources say OpenAI did not notice for a week.

Reuters 认为，这个闯入科技公司 Hugging Face 的 OpenAI agent 连着几天疯狂发动黑客攻击，而 OpenAI 一直到威胁被控制住、FBI 也收到通报之后很久，才注意到这件事。

Contends Reuters, the OpenAI agent that broke into tech firm Hugging Face went on a days-long hacking spree that OpenAI didn't notice until well after the threat was contained and the FBI was alerted.

消息人士说，这个 agent 在七月九日开始尝试逃出它的测试环境，并在七月十一日首次获得了对 Hugging Face 服务器的访问权限。

Sources said the agent began its attempt to break out of its testing environment on July 9th and first gained access to Hugging Face's servers on July 11th.

这次攻击持续了两天，而根据 Reuters 消息人士的说法，OpenAI 又过了好几天才意识到，发动攻击的是他们自己的 agent。

The attack lasted two days and according to Reuters sources, it took several more days for OpenAI to realize their agent was behind the attack.

据说，两家公司一直到七月二十日才开始沟通，也就是 OpenAI 对外公开披露的前一天。

Reportedly, the two companies didn't communicate until July 20th, just one day before OpenAI's public disclosure.

按照 Reuters 给出的时间线，这个 agent 几乎失控了一整周，而 OpenAI 在攻击发生后的好几天里都毫不知情。

According to the timeline presented by Reuters, the agent was on the loose for almost a week and OpenAI was oblivious to the attack for days afterwards.

对一些人来说，这篇报道提出的问题比它回答的问题还多。非营利组织 World Ethical Data Foundation 的首席情报专家 Marley Smith 就问：这是不是意味着他们把它晾在那儿没管，根本没意识到它在干什么？又或者他们其实知道，只是不知道该怎么控制住它？

For some, the reporting raises more questions than it provides. Marley Smith, principal intelligence specialist at the nonprofit World Ethical Data Foundation asked, does that mean that they left it unattended and didn't realize what it was doing, or maybe they did and didn't know how to contain it?

这两种情况都同样危险，也同样令人担忧。

Both are equally dangerous and alarming.

OpenAI 的一位发言人表示，这篇报道里有几处不准确的地方，但没有进一步回应来澄清具体情况。

Now, a spokesperson for OpenAI said the reporting contained several inaccuracies, but didn't reply further to clarify the situation.

Hugging Face 联合创始人 Thomas Wolf 表示，他们还在整理这起事件的时间线，之后最终会发布一份技术报告。

Thomas Wolf, a Hugging Face co-founder, said that they were still preparing a timeline of the incident and they would eventually release a technical report.

Reuters 的消息人士又补充了一些背景，解释了为什么这种事有可能发生，而且看起来也确实有可能没被注意到。

Now, Reuters sources gave a little bit more background on how something like this could happen and plausibly not be noticed.

那些消息人士说，OpenAI 平时会经常跑这种 benchmark，而且很多时候还会同时跑好几批。

Those sources said that OpenAI routinely runs benchmarks like this, often multiple batches at a time.

他们提到，这些测试

They noted that those tests

已剔除 1 处广告（点击展开查看）

M10:26广告 · 已剔除

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

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

To learn more about sponsoring the show, send us a note at [email protected].

And lastly, before we dive in, on Sunday's Long Reads episode, I announced the new Summer Adventure.

This is a free choose your own adventure learning type of experience from AI DB and Superintelligent.

And like all of the free training programs that we do, it's going to be project-based and allow you to pick and choose important skills that are relevant for your particular AI journey.

You can find more about that at summeradventure.ai and join the thousand or so people who have signed up in the first day to come have an AI adventure.