CNN：为什么AI代理不断闯祸越来越失控 | 2026.8.
字幕摘录
| 时间 | 英文 | 中文 |
|---|---|---|
| 0:01 | You heard about this, right? | 你听说了,对吧? |
| 0:04 | An experimental AI model broke loose from its testing environment. | 一个实验性的AI模型脱离了它的测试环境. |
| 0:08 | This kind of sounds like the plot of a sci-fi movie. | 这听起来像科幻电影的情节 |
| 0:10 | It's been called cyber security's Jurassic Park moment. | 这被称为网络安全侏罗纪公园时刻. |
| 0:13 | It is the Sword of Frankenstein moment. | 这是弗兰肯斯坦之剑的时刻。 |
| 0:15 | And launched a days-long targeted cyber attack. | 并发动了为期数天的定向网络攻击. |
| 0:19 | That's a crime, actually. | 其实那是犯罪 |
| 0:20 | Here's how an open AI agent got out, | 这就是一个公开的AI代理是如何出来的, |
| 0:23 | and why it might be tough to stop something like this from happening again. | 以及为什么可能很难阻止这种事再次发生 |
| 0:29 | Okay, so the first thing to know is that | 好吧,所以首先要知道的是, |
展开字幕全文（101 条）
| 序号 | 英文 | 中文 |
|---|---|---|
| 1 | You heard about this, right? | 你听说了,对吧? |
| 2 | An experimental AI model broke loose from its testing environment. | 一个实验性的AI模型脱离了它的测试环境. |
| 3 | This kind of sounds like the plot of a sci-fi movie. | 这听起来像科幻电影的情节 |
| 4 | It's been called cyber security's Jurassic Park moment. | 这被称为网络安全侏罗纪公园时刻. |
| 5 | It is the Sword of Frankenstein moment. | 这是弗兰肯斯坦之剑的时刻。 |
| 6 | And launched a days-long targeted cyber attack. | 并发动了为期数天的定向网络攻击. |
| 7 | That's a crime, actually. | 其实那是犯罪 |
| 8 | Here's how an open AI agent got out, | 这就是一个公开的AI代理是如何出来的, |
| 9 | and why it might be tough to stop something like this from happening again. | 以及为什么可能很难阻止这种事再次发生 |
| 10 | Okay, so the first thing to know is that | 好吧,所以首先要知道的是, |
| 11 | AI companies are testing unreleased AI models and agents all the time. | AI公司一直在测试未释放的AI模型和代理. |
| 12 | And they're testing many of them for raw power, | 他们正在测试其中的许多原始动力, |
| 13 | disabling safeguards installed on publicly released chatbots and agents. | 在公开发布的聊天机和代理机上安装的失效保障措施。 |
| 14 | In early July, open AI researchers created an agent powered by two large language models, | 7月初,开放AI研究者创建了由两个大语言模型驱动的代理器, |
| 15 | and put it inside what's called a sandbox, | 把它放进所谓的沙盒里 |
| 16 | a secure digital environment sealed off from the open internet. | 一个安全的数字环境 关闭了开放的互联网。 |
| 17 | Then they gave it a test called Exploit Gym. | 之后,他们进行了名为"探索健身"的测试. |
| 18 | It's a benchmark cyber security test standardized across the industry | 这是整个行业的标准网络安全测试 |
| 19 | to assess a model's hacking ability, | 评估一个模型的黑客能力, |
| 20 | and it takes even the best AI a few days to complete. | 连最好的AI都要几天才能完成 |
| 21 | Now, the agent had some help, a third-party software, | 现在,代理有一些帮助, 一个第三方软件, |
| 22 | which could go to the outside internet to fetch information and resources the agent might need. | 它可以去外面的互联网 获取信息和资源 代理可能需要。 |
| 23 | Okay, still with me? | 好吧,还和我在一起吗? |
| 24 | Here's where that AI agent got creative. | 这是人工智能特工有创意的地方 |
| 25 | It was like, how can I guarantee I aced this test? | 就像,我怎么能保证 我得到了这个测试? |
| 26 | Here's an idea. | 有个主意 |
| 27 | Instead of taking the test like a good student, why not go find an answer key? | 而不是像一个好学生一样参加考试 为什么不去找个答案? |
| 28 | And I know, this all sounds kind of ridiculous, | 我知道,这听起来有点荒谬, |
| 29 | but don't take my word for it. | 但别听我的话 |
| 30 | It figured out that it could basically cheat on the test. | 它发现它基本上可以欺骗测试。 |
| 31 | See? | 看到没? |
| 32 | But the agent had a problem. | 但经纪人有问题 |
| 33 | There wasn't an answer key in the sandbox. | 沙盒里没有答案 |
| 34 | The only place to find it was out on the open internet. | 唯一能找到它的地方就是在互联网上 |
| 35 | And there was only one way out. | 只有一条出路 |
| 36 | The agent went into villain mode and scoured its helper's buddy's code, | 特工进入了恶棍模式 并搜索了它的助手的好友密码 |
| 37 | found a previously unknown bug in its software, and exploited it | 在其软件中发现了一个先前未知的bug,并对其进行了利用 |
| 38 | to take control of the buddy and Jedi mind trick it to step aside, | 控制好那个好兄弟和绝地的心智 骗它让开 |
| 39 | allowing the model to leave the sandbox and access the open internet. | 允许模型离开沙盒 进入开放的互联网。 |
| 40 | The first thing it did out there was find a clandestine base of operations, | 它首先发现的是秘密行动基地 |
| 41 | a third-party testing sandbox to lurk in for a few days, and took it over. | 第三方测试沙盒 潜伏几天,然后接管。 |
| 42 | As an aside, this is pretty much exactly what Frankenstein's creature did when it escaped the lab. | 顺便说一句,这是弗兰肯斯坦的生物在逃离实验室时所做的。 |
| 43 | Look that up. | 看这个 |
| 44 | So where did it go from there? | 那从那里开始呢? |
| 45 | 由Hugging Face公司主持, | |
| 46 | It's a weird name, I know, but Hugging Face is pretty cool. | 我知道这名字很怪,但"拥抱脸"很酷 |
| 47 | It also hosts collaborative online spaces where independent AI scientists | 它还主机为独立AI科学家合作的在线空间 |
| 48 | post their research for others to learn from. | 把他们的研究寄给其他人学习。 |
| 49 | And it holds a huge archive of AI knowledge and information. | 它拥有大量的AI知识和信息档案. |
| 50 | Think of it as the Library of Alexandria meets Santa's workshop for artificial intelligence. | 把它想成亚历山大图书馆与圣诞老人的人工智能讲习班相遇. |
| 51 | And it's exactly where the answer key is stored. | 而正是答案密钥存放的地方. |
| 52 | Target? | 目标? |
| 53 | Acquired. | 获得. |
| 54 | Over the course of two days, OpenAI's agent volleyed over 17,000 attempts to take over its servers. | 在两天的时间里,OpenAI的特工伏击了超过17,000次试图接管其服务器的尝试. |
| 55 | Hugging Face admins were like, oh snap, and ended up enlisting their own AI model to battle back OpenAI's agent. | 拥抱面部管理员喜欢,哦,snap, 并最终招募自己的AI模型 反击OpenAI的经纪人。 |
| 56 | 直到Hugging Face打电话给OpenAI 看看它是否也被黑了 | |
| 57 | OpenAI的研究人员都喜欢,oops。 | |
| 58 | Their official statement days later, | 他们的正式声明几天后, |
| 59 | quote, we consider this incident to be an unprecedented cyber incident and are responding accordingly. | 引文指出,我们认为这一事件是前所未有的网络事件,并正在做出相应反应。 |
| 60 | Hugging Face wrote, | 笑脸写道: |
| 61 | OK, so now that we've gotten that out of the way, | OK,所以现在,我们已经得到了这一点, |
| 62 | here's why even OpenAI is saying this could happen again. | 所以OpenAI也说这可能会再次发生 |
| 63 | Turns out it wasn't the first time an agent had gone rogue. | 原来不是第一次有特工无赖 |
| 64 | Meta revealed that one of its agents got out, and according to CNN's Hadass Gold, | Meta透露其中一名特工出狱 据CNN的"哈达斯黄金"报道 |
| 65 | Anthropic has also revealed that some of their models were also getting access to the open Internet inside security tests and breaking into other systems. | Anthropic还透露,他们的一些型号也在安全测试内进入开放的互联网,并闯入其他系统. |
| 66 | They didn't know that it happened until they went back and looked at their history, | 他们不知道事情的发生 直到他们回去查看他们的历史 |
| 67 | and they were only prompted to do so after the OpenAI hugging incident. | 在OpenAI拥抱事件之后, |
| 68 | These models are built to get the job done, and they will do anything they can to get that job done. | 这些模型是为了完成这项工作而建造的,他们将竭尽全力完成这项工作。 |
| 69 | They didn't intend for the model to get out and hack a different company system. | 他们并不打算让模型 走出并黑入另一个公司系统。 |
| 70 | But do people in the industry say, cool, like, this is kind of like what we want it to do. | 但行业里的人会说,很酷, 就像我们想要它做的一样。 |
| 71 | Like, these are agents, right? | 他们是特工,对吧? |
| 72 | They're supposed to go out into the world and do things on our behalf. | 他们应该去世界 做一些代表我们的事情。 |
| 73 | The cynic in me is saying, well, this was a great marketing tool for these labs. | 我心中的愤世嫉俗是说, 这对这些实验室来说是一个很好的营销工具。 |
| 74 | Like, look at how good our AI is in that it is making all these decisions on its own. | 比如,看看我们的AI有多好 因为它是 自己做出所有这些决定。 |
| 75 | It can hack into the other systems without anybody knowing. | 它可以在无人知晓的情况下侵入其他系统. |
| 76 | That's crazy. | 这太疯狂了 |
| 77 | But at the same time, there's such a growing call within the industry for rules and regulations over this. | 但与此同时,业界内部也越来越强烈地呼吁对此制定规章制度. |
| 78 | I think the government has a role in setting standards. | 我认为政府在设定标准方面可以发挥作用. |
| 79 | We want to work with the government. | 我们想和政府合作 |
| 80 | There should be some AI regulation. | 应该有一些AI条例。 |
| 81 | But it's not like the federal government hasn't tried to set rules for AI already. | 但联邦政府并没有试图为AI制定规则. |
| 82 | In 2023, then President Joe Biden signed an AI safety executive order that required AI companies to notify | 2023年,时任总统乔·拜登签署AI安全执行令,要求AI公司通知. |
| 83 | and share their testing of frontier models with the government and set the standards for testing. | 并与政府分享对前沿模型的测试,确定测试标准。 |
| 84 | But on his first day back in office, President Trump revoked Biden's AI regulations and replaced it this June | 但特朗普总统在上任的第一天,取消了拜登的AI规定,于今年6月取代. |
| 85 | with a voluntary framework in the service of slashing the bureaucratic constraints | 建立自愿框架,以缓解官僚主义的制约; |
| 86 | that the prior administration placed on America's AI developers. | 之前的行政部门把 美国的AI开发者。 |
| 87 | Under the Trump directive, AI developers can let federal agencies review state-of-the-art models, but they don't have to. | 根据特朗普指令,AI开发者可以让联邦机构审查最新模式,但无需审查. |
| 88 | And the order doesn't mention testing at all. | 而命令完全没有提到测试. |
| 89 | And as of right now, the White House has been careful to emphasize it is all voluntary. | 而现在,白宫一直小心地强调这是自愿的。 |
| 90 | You don't have to submit these models yet. | 你还没有提交这些模型。 |
| 91 | To Hadas, what's happening is more like the rollout of another world-changing technology, the automobile. | 对哈达斯来说,正在发生的事情 更像是推出另一种改变世界的技术,汽车。 |
| 92 | When the car first came out, when the Model T first came out on the road, there were no seat belts, there were no airbags, | 当汽车第一次出来, 当模型T第一次出来 在路上,没有安全带, 没有气囊, |
| 93 | there was no traffic lights. | 没有交通灯。 |
| 94 | It took years for those things to be developed to keep people safe. | 开发这些东西要花好几年才能保证人们的安全。 |
| 95 | The same thing is happening now. | 同样的事情现在正在发生。 |
| 96 | Except, AI is moving much faster than any car ever could. | 除了,AI移动的速度比任何汽车都快得多. |
| 97 | I am bracing for the day when a test model causes real-world harm to real people. | 当一个测试模型对真实的人造成现实世界的伤害时,我正在为这一天吹嘘。 |
| 98 | It hacks a hospital system on accident, gets out, hacks a water utility, didn't intend to. | 它入侵医院系统 意外, 走出, 入侵水的公用事业, 没有打算。 |
| 99 | They'll fix it within a day or something, but it will cause real-world harm, and I'm just waiting for that to happen. | 他们会在一天或什么的时间内修复它,但它会造成现实世界的伤害,我只是在等待这个结果. |
| 100 | And is this the wake-up call to prevent that from ever happening? | 难道这就是防止这种情况发生的警钟吗? |
| 101 | I hope so. | 我希望是这样。 |
该视频共有字幕 101 条。解锁更多字幕为会员功能，请移动到 价格