BuildSpeak每日 builder 文摘
今日归档生词本关于
🎙 播客No Priors· 2026 年 5 月 28 日· 6,814 词 · 约 34 分钟

Building an AI Guardian for Enterprise with Onyx Security CEO Maxim Bar Kogan

SPACE 播放 / 暂停·←→ 上一句 / 下一句
Speaker 100:00 - 00:38
As you're exponentially doing more things with the eyes, you're going to start having really bad actions happen. And we've seen some of that happen lately with agents accidentally publishing code and tokens that they weren't supposed to. Like, definitely, enterprises are starting to realize that risk has grown exponentially and that they don't have any way to stop the adoption. They just now have to do something to reduce the chance of these agent actions being illegitimate or incorrect. But we're allowed to look at a lot of historical data of how these agents have behaved, but enterprises today are not willing to have Anthropic or OpenAI give that historical data because they know these are very data hungry companies that will want to train on that data.
Speaker 100:00 - 00:38
随着你通过“眼睛”以指数级方式做越来越多的事情,真的会开始出现非常糟糕的动作。最近我们已经看到了一些这类情况:agent 会意外发布它们本不该发布的代码和 token。可以说,enterprise(企业)现在开始意识到,风险已经指数级增长,而且他们没有任何办法阻止这种采用。现在他们只能想办法降低这些 agent 动作不合法或不正确的概率。我们当然可以去看大量关于这些 agent 过去如何行为的历史数据,但如今企业并不愿意让 Anthropic 或 OpenAI 获取这些历史数据,因为他们知道这些公司非常“饥渴”于数据,而且会想用这些数据来训练模型。
Speaker 200:44 - 01:06
Hi, listeners. Welcome back to KnowPriars. Today, I'm here with Maxim Bar Kogan, the co founder and CEO of Onyx Security, an Israel based startup of researchers, mathematicians, engineers building agents to watch the AI agents. We talk about specialized model training, mythos, alignment research, and the Israeli ecosystem in security and now AI. Welcome.
Speaker 200:44 - 01:06
各位听众,大家好。欢迎回到 KnowPriars。今天和我一起的是 Maxim Bar Kogan,他是 Onyx Security 的联合创始人兼 CEO。Onyx Security 是一家位于 Israel 的 startup(初创公司),由研究人员、数学家和工程师组成,专门构建用来监视 AI agents 的 agents。我们聊了专用模型训练、mythos、alignment research,以及 Israel 在 security 以及如今 AI 领域的生态。欢迎。
Speaker 201:06 - 01:08
Maxim, thanks so much for doing this.
Speaker 201:06 - 01:08
Maxim,非常感谢你来参加。
Speaker 101:08 - 01:10
Thank you. Pleasure to be here.
Speaker 101:08 - 01:10
谢谢,很高兴来到这里。
Speaker 201:10 - 01:40
Everyone is much more concerned about security and the impact of AI on security than they were certainly a few months ago. The consensus risk story two years ago when you started the company was basically like DLP for chatbots. Like, what are employees putting into ChatGPT? Now we have clearly something that is not quite panic, but close to market wide panic. How did you decide to bet on agent actions when you started?
Speaker 201:10 - 01:40
现在,大家对 security,以及 AI 对 security 的影响,显然比几个月前担心得多得多。两年前你们创办公司时,关于风险的主流叙事基本还是“给 chatbot 做 DLP”——比如员工往 ChatGPT 里输入了什么。现在我们显然面对的局面还不算彻底恐慌,但已经接近全市场范围的恐慌了。你们当初创业时,是怎么决定把赌注押在 agent actions 上的?
Speaker 101:40 - 01:50
Look, I think for us, the pivotal point was AutoGPT. I think AutoGPT kind of let everyone's imagination, including ours, run wild because it was a
Speaker 101:40 - 01:50
你看,我觉得对我们来说,关键转折点是 AutoGPT。我觉得 AutoGPT 某种程度上让所有人——也包括我们——的想象力彻底放飞了,因为它是一个
Speaker 201:50 - 01:52
Can you remind listeners what that was?
Speaker 201:50 - 01:52
你能提醒一下听众那是什么吗?
Speaker 101:52 - 02:26
Sure. So AutoGPT, and I'm sorry if I don't know the guy behind it, but a huge fan, they created the first, as far as I know, the first really autonomous agent running on LLMs, right? So, an agent that would let LLM not generate text, but decide what to do and then give that agent an API access to do that thing, a tool to do it. And then we'd do that in a loop. So, it basically, in theory, could let agents do very complicated things, anything a person could do on a computer.
Speaker 101:52 - 02:26
当然。AutoGPT——如果我不知道背后那个人是谁,还请见谅,但我是他的超级粉丝——据我所知,他们做出了第一个真正运行在 LLMs 之上的自主 agent。也就是说,这种 agent 不只是让 LLM 生成文本,而是让它决定要做什么,然后再给这个 agent 相应的 API 访问权限去执行那件事,也就是给它一个完成该事的工具。接着它会在一个循环里反复这么做。所以从理论上说,它基本可以让 agents 去完成非常复杂的事情,任何一个人能在电脑上完成的事都可以。
Speaker 102:26 - 03:09
Now, granted it didn't work that well, it was too early, the models were not good enough, GPT-four was not good enough, but I think it did give everyone a glimpse into the future of what if the models were good enough, and then basically using that same structure, we could have very capable agents doing stuff for us. I think that was, in many ways, cloud code today is not dissimilar to AutoGPT back then. I think they were a bit early again, before the models were ready, but the concept was right. And the thought that sticked with me was I was very eye pilled even back then. So, I was thinking, oh my god, models are gonna be way smarter than us.
Speaker 102:26 - 03:09
当然,先说明一下,当时它并没有真正运作得那么好;时机太早了,models(模型)也还不够好,GPT-four 也还不够好。但我觉得,它确实让所有人瞥见了未来的一角:如果 models 足够好了,会怎样?然后基本上沿用同样的结构,我们就能拥有非常强大的 agents(智能体)替我们做事。我觉得从很多方面看,今天的 cloud code 和当年的 AutoGPT 并没有那么不同。我认为他们又一次有点太早了,早于 models 准备好的时间,但概念是对的。而当时一直留在我脑海里的想法是,我那时就已经非常 eye pilled 了。所以我当时在想,天哪,models 会变得比我们聪明得多。
Speaker 103:09 - 03:31
When that happens, how do we oversee these very smart agents that are smarter than us, they're very capable? How are going to feel easy about them doing stuff for us, especially when they start managing really important stuff. One day they're managing your water supply and your electricity, your power grid. Right? How do you control them?
Speaker 103:09 - 03:31
当这种情况发生时,我们该如何监督这些比我们更聪明、能力又非常强的 agents?我们怎么才能安心让它们替我们做事,尤其是当它们开始管理真正重要的事情时?总有一天,它们会管理你的供水、你的电力、你的 power grid(电网)。对吧?你要怎么控制它们?
Speaker 103:31 - 03:50
And that was like the thing I was kind of obsessed about that thought. I was also too early. So, I think at the time, enterprises were not using any agents. There were hardly any agents out there. And talking with a lot of security buyers at the time, they were like, oh, dude, you're way too early.
Speaker 103:31 - 03:50
这几乎就是那个让我有点着迷、反复琢磨的核心问题。只是我也同样太早了。所以我觉得,在那个时候,enterprise(企业)还根本没有在使用任何 agents,市面上也几乎没有什么 agents。那时我和很多 security(安全)采购方聊,他们都说,哦,兄弟,你这也太超前了。
Speaker 103:50 - 03:52
Like, this is not something that's gonna happen. I asked
Speaker 103:50 - 03:52
就像,这种事根本不会发生。我还问过
Speaker 203:52 - 03:56
you the same question. I said, is anyone going to do this before you run out of money?
Speaker 203:52 - 03:56
你同样的问题。我说,在你把钱烧光之前,真的会有人做这个吗?
Speaker 103:58 - 04:46
I think there was a good chance that I would've run out of money before because I think you were right. Like, I think there was an element of chance here, but then I think the market did happen. So, we had suddenly reasoning models that could do long horizon tasks. We had Cloud Code, which became the really first widely used autonomous agent, and then we had Cowork and OpenCLO. And I think we're starting to see now that these types of agents that are very autonomous, even though everyone was afraid to build them, everyone started building these low code platforms that were much more limited, much more based on connectors, those platforms ended up being quite limited, so we didn't get the productivity gains from those limited platforms.
Speaker 103:58 - 04:46
我觉得我在那之前很有可能真的会把钱花光,因为我认为你当时是对的。就是说,我觉得这里面确实有运气成分。但后来我认为市场确实发生了。于是我们突然有了 reasoning models(推理模型),它们可以处理 long horizon tasks(长周期任务)。我们有了 Cloud Code,它成了第一个被广泛使用的 autonomous agent(自主智能体);然后我们又有了 Cowork 和 OpenCLO。我觉得我们现在开始看到,这类高度 autonomous 的 agents,尽管之前所有人都害怕去构建它们,大家还是都转而去做那些 low code(低代码)平台,那些平台受限得多,更依赖 connectors(连接器);结果这些平台最终都相当受限,所以我们并没有从这些受限的平台中获得多少生产力提升。
Speaker 104:46 - 05:16
But when we started getting the crazy benefits from these very unleashed agents that could do everything, that had much less controls baked into them, and even very large enterprises decided they're going to adopt it, you know, like Anthropix revenue is coming from enterprises that are paying for cloud code to do a lot of the work that developers used to do. That was a bit of kind of how we started, and we definitely were in luck that very autonomous agents appeared before it was too late.
Speaker 104:46 - 05:16
但当我们开始从这些几乎“放开手脚”的 agents 身上获得惊人的收益时——它们几乎什么都能做,内置在其中的 controls(控制机制)也少得多——甚至非常大型的 enterprise 都决定采用它。你知道的,比如 Anthropix 的 revenue(营收)就来自这些 enterprise:它们付费让 cloud code 去完成许多原本由开发者来做的工作。这大概就是我们起步的方式,而我们也确实很幸运:高度 autonomous 的 agents 在一切还没来不及之前就出现了。
Speaker 205:17 - 05:36
So can you describe a little bit just because it's, I think, both close to impossible and then very useful in this period of AI to think about what is deployment right now and then, you know, what's changing about capability? What's the one liner on what the Onyxx product does today? And then, like, how you think about long term vision?
Speaker 205:17 - 05:36
所以你能不能稍微描述一下——因为我觉得在当前这个 AI 阶段,这件事既几乎不可能说清,又非常有用——所谓 deployment(部署)现在到底意味着什么,以及 capability(能力)正在发生什么变化?如果用一句话概括,Onyxx 这个产品今天到底是做什么的?然后,再说说你如何看待它的长期愿景?
Speaker 105:36 - 06:41
Today, like, Onyxx is really does two things. Number one is we train models and build agents that can oversee other agents. And the goal of that is to say, okay, we need someone to be able to tell that all of these actions that are now happening by these AIs that we're adopting are legitimate because the number of these actions is growing exponentially. And so, things that we thought might be useful in the past, like a human in the loop, now that you're going to have 100x, 1000x, a million x of these actions, that's not going to work. And then we take that capability and we basically productize it in a product that we call the AI control plane or the secure AI control plane, where we come to Enterprise and say, hey, let's find all of your AIs and autonomous agents and hook them up to Onyx to this system where we can oversee what your AIs are doing, so that you don't run into the risk of, as you're exponentially doing more things with AIs, you're going to start having really bad actions happen.
Speaker 105:36 - 06:41
如今,Onyxx 实际上主要做两件事。第一,我们训练模型并构建能够监督其他 agent 的 agent。其目标是:我们需要有人——或者说某种机制——能够判断,现在这些被我们采用的 AI 所执行的所有动作是否合规,因为这类动作的数量正在指数级增长。因此,过去我们以为可能有用的做法,比如 human in the loop(人类参与闭环),在未来这些动作变成 100 倍、1000 倍、100 万倍之后,就不再可行了。接着,我们把这种能力产品化,做成一个我们称为 AI control plane,或者 secure AI control plane 的产品。我们会对 Enterprise 说,来吧,让我们找出你们所有的 AI 和 autonomous agents,把它们接入 Onyx,接入这个系统,这样我们就能监督你们的 AI 在做什么,避免你们在用 AI 指数级扩大工作量时,开始出现一些非常糟糕的行为。
Speaker 106:41 - 07:11
And we've seen some of that happen lately with downtimes that were caused by agents doing their own thing, agents accidentally publishing code and tokens that they weren't supposed to, and so on. So, definitely, enterprises are starting to realize that that risk has grown exponentially and that they don't have any way to stop the adoption. So, they just now have to do something to reduce the chance of these agent actions being illegitimate or incorrect.
Speaker 106:41 - 07:11
而且我们最近已经看到一些这样的情况发生了:比如 agent 各行其是导致的 downtime(停机),agent 意外发布了本不该发布的代码和 token,等等。所以很明显,enterprise 现在开始意识到,这种风险也在指数级增长,而他们又没有办法阻止 AI 的采用。因此,他们现在只能采取措施,尽量降低这些 agent 行为不合规或不正确的概率。
Speaker 207:11 - 07:41
Yeah. I think one of the core reasons, obviously, the foundation model labs are going after code is because it is very powerful in general and can do, you know, in theory, all things software can over time. The flip side of that is it can do all things software can. Right? And so joyously am already in the camp of a, having been over permissive with my agents such that it deleted data permanently and caused rework.
Speaker 207:11 - 07:41
是的。我认为其中一个核心原因,显然是 foundation model labs 正在攻代码这个方向,因为代码总体上非常强大,而且理论上,随着时间推移,它可以完成 software 能做的所有事情。问题的另一面在于,它也确实可以完成 software 能做的所有事情,对吧?所以我自己其实已经属于那种——曾经给我的 agent 放权过多,结果它永久删除了数据,还造成了返工——的阵营了。
Speaker 207:41 - 07:41
So I'm like,
Speaker 207:41 - 07:41
所以我就想,
Speaker 107:41 - 07:42
oh, okay.
Speaker 107:41 - 07:42
哦,明白了。
Speaker 207:42 - 08:08
I think I need some guardian spirits around it. Given your deployments today and talking to large enterprises, what is the state of deployment? Right? Like, much do you see that's within these more scoped, like, studio like platforms versus, you know, free riding coding agents? How much are you actually seeing in large enterprises and in different sectors?
Speaker 207:42 - 08:08
我觉得我需要在它周围放一些“guardian spirits(守护机制)”。结合你们现在的部署情况,以及你们和大型 enterprise 的交流,目前的部署状态到底是什么样?比如说,你们看到的更多是在这些范围更受限、类似 studio 的平台里,还是那种更自由发挥的 coding agents?在大型 enterprise 以及不同垂直行业里,你们实际看到的比例分别有多少?
Speaker 108:09 - 08:34
Yeah. So, I think right now, in our typical enterprise, we break it down to three categories. So, we break it down to various SaaS platforms that are typically more low code, where people build agents in this drag and drop way, and they're not really autonomous agents. I would think of them more as AI automations. Then there are first party agents.
Speaker 108:09 - 08:34
是的。所以我觉得,目前在我们接触的典型 enterprise 里,我们会把它分成三类。第一类是各种 SaaS 平台,通常更偏 low code,人们用拖拽式的方法在上面构建 agent,但它们其实不算真正的 autonomous agents。我更愿意把它们看成 AI automations。第二类是 first party agents。
Speaker 108:34 - 09:22
People are building in their cloud, potentially because it's an application they want inside the company, or even a product they're planning to release to the customers that is agentic. And then the third category is very autonomous coding agents and assistants. Of these categories, I would say roughly at this point, over 50% is the autonomous coding agents and assistants in the average enterprise. Then probably 45% is those low code automations, and the last 2% are really the first party ones that they're building themselves, because obviously it's much harder to build effective agents, and it's much easier to adopt agents off the shelf or build them with low code. And that's what we're seeing.
Speaker 108:34 - 09:22
也就是人们在自己的 cloud 里构建的 agent,可能因为这是他们希望在公司内部使用的应用,或者甚至是他们计划发布给客户、并且带有 agentic 特性的产品。第三类则是高度 autonomous 的 coding agents 和 assistants。在这些类别中,我会说,现阶段平均来看,超过 50% 是 autonomous coding agents 和 assistants;大概 45% 是那些 low code automations;最后大约 2% 才是真正由他们自己构建的 first party agent,因为显然,构建有效的 agent 要困难得多,而直接采用现成的 agent 或者用 low code 去搭建则容易得多。我们现在看到的情况就是这样。
Speaker 109:22 - 09:58
We do think that the automations are also the fastest growing category. So, it used to be that only developers would see cloud code growing like fire in our customer base, and now we're seeing a cloud co worker growing even faster. We're starting to see, to our own surprise actually, people adopting OpenCloud as a legitimate sanction tool in the company because the CEO is very driven to adopt AI. So, I think that today, autonomous Azure is by far the fastest growing category, today typically comes without any controls.
Speaker 109:22 - 09:58
我们确实认为,automations 也是增长最快的类别。过去,只有 developers 会看到 cloud code 在我们的客户群中像野火一样增长,而现在我们看到一种 cloud co worker 增长得更快。实际上,连我们自己都有些惊讶:因为 CEO 非常积极地推动采用 AI,人们开始把 OpenCloud 当作公司里一种正式获批的工具来采用。所以我认为,今天 autonomous Azure 绝对是增长最快的类别,而且目前通常是在几乎没有任何控制措施的情况下被引入的。
Speaker 209:58 - 10:23
So enterprises already buy, let's say, 100,000,000,000 of security today. They have lots of different protections at the endpoint and network and cloud and identity domains. What's relevant here for securing agents, or is none of it? Like, how do you think about the existing protection set?
Speaker 209:58 - 10:23
比如说,企业今天已经购买了 100,000,000,000 规模的安全产品。他们在 endpoint、network、cloud 和 identity 等不同领域都有很多不同的防护措施。这里真正相关的问题是:这些现有东西对保护 agents(智能体)有帮助吗,还是说基本都没用?你会如何看待这套现有的防护体系?
Speaker 110:24 - 11:01
Security is always a space where you have some overlap between different tooling. In this case, and you have the concept of defensive debt as well. So, you want to have defenses at different levels of your technology stack to solve the problem. And that said, I think in this space, a lot of enterprises are kind of helpless because I'll take an example, the identity approach. Traditionally, if we have a software system that's running in our company, our first and most important control will be to limit what permission it has.
Speaker 110:24 - 11:01
安全领域里,不同工具之间总是会有一些重叠。在这个场景下也是如此,而且你还会面对 defensive debt(防御债务)这个概念。所以,你会希望在技术栈的不同层级都部署防御措施来解决问题。话虽如此,我认为在这个领域里,很多企业其实有点束手无策。我举个例子,比如 identity 这条思路。传统上,如果公司里运行着一个软件系统,我们第一也是最重要的控制手段,就是限制它拥有哪些权限。
Speaker 111:01 - 11:55
And then no matter what, even if it goes wrong, even if it's compromised, it can't typically do stuff that it was originally allowed to do. But with these autonomous AIs, with these assistants, with these coding agents, we kind of want them to have our permissions because we want to tell Cloud Code to do something or Cloud Code work to do something, and we want to then go have lunch, we want to come back and see that it's done. And we want to give it so many diverse tasks as well that we kind of can't find the right set of permissions to do. So, suddenly our identity security software is not very useful. Then if you think about endpoint security or API security, if we tell our Cloud Code that we want to recreate a database and it should delete it and recreate it, that's great.
Speaker 111:01 - 11:55
这样一来,无论发生什么情况,即使它出错了,即使它被攻破了,通常它也做不了那些原本没有被允许做的事。但对于这些 autonomous AIs(自主 AI)、这些 assistants(助手)、这些 coding agents(编码智能体),我们某种程度上又希望它们拥有我们的权限,因为我们想告诉 Cloud Code 去做某件事,或者让 Cloud Code work 去做某件事,然后我们就可以去吃午饭,回来时希望它已经完成了。而且我们还希望它承担很多种不同的任务,以至于我们几乎找不到一组合适的权限来覆盖这些需求。于是,我们原有的 identity security software(身份安全软件)突然就不那么有用了。再比如 endpoint security 或 API security,如果我们告诉 Cloud Code,我们想重建一个数据库,它应该先删除再重建,那当然很好。
Speaker 111:55 - 12:19
That's gonna save our DevOps team and our platform teams a lot of time. It's a great benefit of Cloud Code. But if Cloud Code is working on an unrelated task and suddenly thinks that maybe the right thing to do is to delete our database and recreate it. Maybe we don't want that to happen. And unfortunately, our endpoint providers or API security tools, they don't know what cloud was thinking, why is it doing what it's doing, right?
Speaker 111:55 - 12:19
这会为我们的 DevOps 团队和 platform 团队节省大量时间。这正是 Cloud Code 的一个很大价值。但如果 Cloud Code 明明在处理一个无关任务,却突然觉得“也许正确的做法是删除我们的数据库并重新创建它”,那这种事我们可能并不希望发生。遗憾的是,我们的 endpoint providers 或 API security tools 并不知道 cloud 当时在想什么,也不知道它为什么要这么做,对吧?
Speaker 112:20 - 12:44
So, a lot of these existing tools, they don't have the context to understand what these very flexible, unpredictable systems are doing. If you're not building some kind of controls that are built for these systems, then you're either going to end up limiting them a lot, making them almost much less useful to the enterprise, or you're going to miss a lot of pretty dangerous things that they might be doing.
Speaker 112:20 - 12:44
所以,很多现有工具都缺乏足够的上下文,无法理解这些高度灵活、不可预测的系统到底在做什么。如果你不构建某种专门面向这类系统设计的控制机制,那你最后要么会把它们限制得很死,让它们对企业几乎失去大部分价值;要么就会漏掉很多它们可能正在做的、相当危险的事情。
Speaker 212:45 - 13:02
As somebody who has worked in security for a long time, my first very traditional instinct on a problem like this is like, that sounds like a problem for a proxy with a policy engine. We make some rules and make the rules smarter. Like, why doesn't that work, or did you try it?
Speaker 212:45 - 13:02
作为一个在安全领域工作了很久的人,我面对这类问题时,第一反应是非常传统的:这听起来像是一个应该交给 proxy(代理)加上 policy engine(策略引擎)去解决的问题。我们制定一些规则,再把规则做得更智能一些。为什么这条路行不通,还是说你们已经试过了?
Speaker 113:03 - 13:36
There are a few things that, I mean, proxy's integration method, I would say. So, there are some AI systems where, like, you would want to integrate with a proxy if that's the easiest way to do it. But number one, there's a lot of systems where that's just not viable technically, because AI today runs on the cloud, on someone else's infrastructure, on your endpoint, and just proxy is not always an option. And the second thing is the question, okay, great, you're proxying, so you're seeing the data. You're seeing it, but that's not the hard problem.
Speaker 113:03 - 13:36
这里面有几件事,我会说首先是 proxy 的 integration method(集成方式)问题。的确,有些 AI 系统里,如果通过 proxy 集成是最简单的方法,那你会想这么做。但第一,有很多系统在技术上根本不可行,因为今天的 AI 既可能运行在 cloud 上、运行在别人的基础设施上,也可能运行在你的 endpoint 上,所以 proxy 并不总是一个可选项。第二个问题是:好,就算你做了 proxying(代理转发),你因此看到了数据。你是看到了,但这并不是最难的问题。
Speaker 113:36 - 14:05
The hard problem is understanding what I should do now. Turns out that in the case of AI systems, that is the hard question. Like, what is the engine that needs to underwrite these different actions and say if they're okay or not? Because we need to be able to understand what another AI system is thinking, what is it planning to do, and then have our own opinion on that. And consider we're trying to understand some of the smartest models in the world are doing the right thing.
Speaker 113:36 - 14:05
真正难的问题,是理解我现在应该做什么。结果发现,在 AI system(人工智能系统)的场景里,这恰恰就是最难的问题。比如说,到底需要一个怎样的 engine(引擎)来为这些不同的行动提供底层支撑,并判断它们是否可以接受?因为我们需要能够理解另一个 AI system 在想什么、它打算做什么,然后再对此形成我们自己的判断。再想想看,我们现在试图理解的,还是世界上一些最聪明的 model(模型)是否在做正确的事。
Speaker 114:05 - 14:11
So, who are we to do it? How are we going to do it correctly? Right? And so, that turns out to be a really difficult technical question.
Speaker 114:05 - 14:11
所以,我们凭什么来做这件事?我们又要怎样才能把它做对?对吧?而这最终就变成了一个非常困难的技术问题。
Speaker 214:11 - 14:18
Part of the solution for Onyx has been training its own models. Like, what can you say about that?
Speaker 214:11 - 14:18
对 Onyx 来说,解决方案的一部分是训练自己的 model(模型)。这一点你可以谈谈吗?
Speaker 114:18 - 14:52
If you try today, let's say we're trying to build a solution to oversee and kind of control how other agents are operating. Maybe the first thing a lot of our listeners might think is say, well, I'll just ask Cloud Code to do it. In a sense, they would be right because Cloud Code is great. And maybe we can ask it to spawn a version of itself for every agent that we have and kind of keep monitoring everything that agent is starting to do. And if you think that there's a problem, intervene.
Speaker 114:18 - 14:52
如果放在今天来尝试,比方说我们想构建一个 solution(解决方案),去监督并某种程度上控制其他 agent 的运行方式。也许很多听众首先会想到的是:那我直接让 Cloud Code 来做不就行了。某种意义上,他们也没错,因为 Cloud Code 很强。也许我们可以让它为我们拥有的每个 agent 都 spawn(生成)一个自己的版本,然后持续监控那个 agent 开始做的每一件事。一旦你觉得有问题,就进行干预。
Speaker 114:52 - 15:37
So, that approach, obviously it's pretty naive, and there are some ways in which it totally fails, we could talk about, but it has some merit to it. So, it does seem intuitive that it's a good idea to have capable agents reviewing what other agents are doing, same as we have capable humans reviewing what other humans are doing. But then the problems that you're gonna run into is how do I make this work from cost, latency, reliability perspective? Because if I need to run an agent for every agent you're running as your security vendor, you're gonna be paying for me more than you're paying for your AI, right? So, it's pretty much a dealbreaker.
Speaker 114:52 - 15:37
这种方法显然相当 naive(朴素),而且它在某些方面会彻底失效,这些我们可以展开讲,但它也确实有一定道理。直觉上看,让有能力的 agent 去审查其他 agent 在做什么,确实是个好主意,就像我们会让有能力的人去审查其他人在做什么一样。但接下来你会遇到的问题是:从 cost(成本)、latency(延迟)、reliability(可靠性)这些角度看,我要怎样才能让它真正可行?因为如果作为你的 security vendor(安全供应商),我需要为你运行的每一个 agent 再运行一个 agent,那你付给我的钱可能会比你为自己的 AI 付的钱还多,对吧?所以这基本上就是个 dealbreaker(无法接受的问题)。
Speaker 115:37 - 16:16
And also, it's going be so slow, so you're not going be happy with whatever latency you're going to get. And so, the challenge then becomes how do I know what are the times where I need to interject with these smart agents to look at what's happening? And that's when actually what you want to do is you want to train very smart models that are, actually let me correct with that, very not smart models, but models that are just good at one thing. They're very small. They almost can't do anything else other than be able to say, should I have a smarter agent look at this?
Speaker 115:37 - 16:16
而且它还会非常慢,所以你肯定不会满意最终得到的 latency(延迟)。于是挑战就变成了:我怎么知道,在哪些时刻我需要插入这些聪明的 agent,去查看正在发生什么?这时候你真正想做的,其实是训练一些非常聪明的 model(模型)——不过让我纠正一下,不是非常聪明的模型,而是只擅长一件事的模型。它们非常小,几乎什么别的事都做不了,唯一擅长的就是判断:这件事是否应该交给一个更聪明的 agent 来看一眼?
Speaker 116:16 - 17:14
And if you manage to bake in that intuition into those small models well, in the sense that they don't miss a lot of stuff and they don't call that other agent too much, then you can get to a really good balance of we're very performant, we have smart agents overseeing things when needed, our costs are low and our latency is low. And then that becomes the challenge because you need to make sure that as the frontier models get smarter and the hardnesses become more evolved, you need to be able to have models that are on your side that are small and effective and continuously being able to say, now is the time. This is the action where I think someone should take a closer look. And that's why we want to experience models for this purpose, and it's you know, most of the hard things that we're doing in this space.
Speaker 116:16 - 17:14
如果你能把这种直觉很好地 bake in(固化)到那些小模型里,也就是说,它们不会漏掉太多东西,也不会过于频繁地去调用另一个 agent,那你就能实现一个非常好的平衡:我们的 performance(性能)很高,需要时有聪明的 agent 负责监督,同时成本低、延迟也低。接着这又带来了新的挑战,因为随着 frontier models(前沿模型)越来越聪明,问题的难度也会不断演化,你就必须拥有站在你这边的模型——它们小而高效,并且能够持续判断:现在就是该介入的时候了,这个 action(动作)是我认为应该让人更仔细看一看的。也正因为如此,我们才想为这个目的去打磨模型,而这也是我们在这个领域所做的大多数困难工作的核心。
Speaker 217:14 - 17:40
Yeah. You and I actually both love to play blitz chess, and I look at Guardian as a system that's a little bit analogous. Like, it's not clear either of us is going to be competitive with Magnus in a real game. But if the if you play if you play enough times with the right data and all you have to do is make intuitive decisions under time pressure very, very quickly, it's actually a different game. Right?
Speaker 217:14 - 17:40
对。你我其实都很喜欢下 blitz chess(快棋),而我看 Guardian 这个系统时,会觉得它有点类似这种情况。比如说,在一盘真正的对局里,我们俩谁都未必能和 Magnus 竞争,这一点并不明确。但如果你在拥有合适数据的前提下进行足够多次训练,而且你要做的只是要在时间压力下非常、非常快地做出直觉判断,那它其实就变成了另一种游戏。对吧?
Speaker 217:40 - 17:43
And do you think do you think that makes sense, or am I reaching here?
Speaker 217:40 - 17:43
你觉得这说得通吗,还是我这是在过度联想?
Speaker 117:43 - 18:17
Yeah. I actually I didn't think about it, but, yeah, there's a lot of analogies because I think if you look at top chess players in the world, like, most of the moves that they make are intuitive. They don't calculate forward. They've seen so much games and they've played so much games that they already have a good sense of what is their right move and that they're not taking too much risk here by taking this move without calculating. And then if you look at those games, every once in a while, they do stop for suddenly a really long period of time to just calculate forward a lot of options because they know this is a critical move in the game.
Speaker 117:43 - 18:17
对。其实我之前没这么想过,不过,确实有很多类比,因为如果你看世界顶尖的 chess(国际象棋)选手,比如说,他们下的大多数步其实都是直觉性的。他们不会一路往前计算。他们看过太多对局,也下过太多对局,所以他们已经对什么是正确的一步有很强的感觉,也知道在不做计算的情况下走这一步并没有承担太大风险。然后如果你去看那些对局,时不时他们也会突然停下来很长一段时间,去大量向前计算各种选择,因为他们知道这是棋局中的关键一步。
Speaker 118:17 - 18:37
There's risk, You need to think through what you're doing and you need to decide correctly. I think that's very similar. It's the efficient way to run computation, right? You don't want to spend too much intelligence where you don't have to, and you want to spend a lot of intelligence, overwhelmingly a lot, in situations where there's high risk.
Speaker 118:17 - 18:37
这里是有风险的。你需要把自己在做什么想清楚,而且你需要做出正确决策。我觉得这非常相似。这其实是运行 computation(计算)的一种高效方式,对吧?你不想在不必要的地方投入过多 intelligence(智能/算力),而你会希望在高风险情境下投入大量的 intelligence,压倒性地大量投入。
Speaker 218:37 - 19:08
You guys are a team mostly based in Israel today. I think the world has accepted that there is a cohort of amazing Israeli security talent that comes out of, you know, the military and offensive security and then, you know, repeat entrepreneurs like you guys. I think the DNA at Onyx is a little bit different here. Your co founder Gil came out of building synthetic data and working at NVIDIA. Like, What would you characterize the talent that Onyx is particularly good at?
Speaker 218:37 - 19:08
你们团队现在主要是在 Israel。我觉得全世界都已经接受这样一个事实:Israel 有一批非常出色的 security(安全)人才,来自军方、offensive security(攻防安全),以及像你们这样的连续创业者。我觉得 Onyx 的 DNA 在这方面又有点不一样。你的联合创始人 Gil 是做 synthetic data(合成数据)出身,也在 NVIDIA 工作过。那么,你会怎么概括 Onyx 特别擅长的人才类型?
Speaker 219:08 - 19:14
And then, are people actually training interesting frontier models in Israel now?
Speaker 219:08 - 19:14
还有,现在真的有人在 Israel 训练有意思的 frontier models(前沿模型)吗?
Speaker 119:14 - 20:01
So, of I think Israel started maybe a bit late in the game, but it's catching up quickly. So, I think there's now amazing companies in Israel building world models, building AI infrastructure that's top of its glass, building chips. So, I think Israel, in general, is becoming very strong in AI, and we're proud to be a part of that movement. And I think you're right, our company has a very mixed DNA between cyber and AI, which kind of reflects mine's and Gil's backgrounds. Most of the people in our company, most of our research engineering come from a unit in the Israeli intelligence where we actually deal with math and cyber and the intersection thereof.
Speaker 119:14 - 20:01
所以,我觉得 Israel 可能起步稍微晚了一点,但它追赶得很快。所以我认为,现在 Israel 已经有一些很棒的公司在构建 world models(世界模型)、构建顶级的 AI infrastructure(AI 基础设施)、以及造芯片。所以我觉得,总体上 Israel 正在变得非常强于 AI,而我们也很自豪能成为这股浪潮的一部分。而且我觉得你说得对,我们公司的 DNA 确实是 cyber(网络安全)和 AI 的混合体,这某种程度上也反映了我和 Gil 的背景。我们公司里的大多数人,尤其是大多数 research engineering(研究工程)人员,都来自 Israel intelligence(以色列情报部门)的一个单位,在那里我们实际处理的是数学、cyber,以及两者的交叉领域。
Speaker 120:02 - 20:26
And so, I think it is also reflected in kind of the type of talent that we bring in. I think it's important for a few reasons. The first and foremost is that we want to be more than just a security company long term. We think that to solve this problem well, it's going to require deep AI expertise, but then the problem is not just cybersecurity. The problem is how do we control advanced AI long term?
Speaker 120:02 - 20:26
所以,我觉得这也反映在我们吸纳的人才类型上。我认为这之所以重要,有几个原因。首先也是最重要的一点是,从长期来看,我们想做的不只是一个 security company(安全公司)。我们认为,要把这个问题真正解决好,将需要非常深厚的 AI 专业能力,但这个问题又不只是 cybersecurity(网络安全)问题。真正的问题是:从长期来看,我们该如何控制 advanced AI(高级 AI)?
Speaker 120:26 - 21:21
And that problem, even if you just forget about enterprise security and the different gaps in various controls that they currently have, first principle is that problem just sounds very important to me. So, I think it will be crucially important. If you have AI companies that are $10,000,000,000,000 companies, we think you want a company that is not the vendor of the AI itself to oversee and help you control what AI is doing. And we think that's an opening that's a $100,000,000,000 plus opening for a really important company. And then if you think about what it's gonna take to control advanced AI long term, then we're just scratching the surface because long term, you're going to have to also understand much better what models are thinking, what's happening on the internals of these models as they're operating.
Speaker 120:26 - 21:21
而这个问题,即便你先不去考虑 enterprise security(企业安全),也不去考虑他们当前各种控制手段中的不同缺口,从第一性原理来看,这本身就是一个对我而言非常重要的问题。所以,我认为它会变得至关重要。如果未来出现的是市值 $10,000,000,000,000 的 AI 公司,我们认为你会希望有一家公司——不是 AI 本身的 vendor(供应商)——来监督并帮助你控制 AI 在做什么。我们认为,这会是一个价值超过 $100,000,000,000 的机会窗口,足以诞生一家非常重要的公司。然后,如果你去想,从长期来看,要控制 advanced AI 需要什么,那么我们现在其实才刚刚触及表面,因为从长期看,你还必须更好地理解 models(模型)在“思考”什么,这些模型在运行时其内部到底发生了什么。
Speaker 121:22 - 21:24
That's also a lot of where our research is focused.
Speaker 121:22 - 21:24
这也是我们研究重点所在的很大一部分。
Speaker 221:25 - 21:42
So, the industry is quite divided on this issue. I mean, amongst the people who think about whether or not mechanistic interpretability or research into better understanding models is possible? Like, that's a question. So, it's something you believe in.
Speaker 221:25 - 21:42
所以,行业内部在这个问题上的分歧相当大。我的意思是,在那些思考“mechanistic interpretability(机制可解释性)”是否可行、或者研究是否有可能更好地理解模型的人当中,这本身就是一个问题。所以,这是你所相信的一件事。
Speaker 121:43 - 22:25
We believe that there's been a lot of strong progress in that direction. We believe that understanding the internal weights and activations, what is the internal structure, the mathematical structure of these systems, is gonna be at least part of the solution. And in many ways, we think that, and this is maybe, we'll only know when we get there, but we think that for our level of intelligence, it's kind of difficult to understand very quickly what is the internal structure of a large language model. What is the internal structure of way Our it's an
Speaker 121:43 - 22:25
我们相信,朝这个方向已经取得了很多强有力的进展。我们相信,理解内部的 weights(权重)和 activations(激活),理解这些系统的内部结构、数学结构,至少会是解决方案的一部分。而且在很多方面,我们认为——当然这也许只有到那时才能真正知道——以我们这个 intelligence(智能)水平来说,要非常快速地理解 large language model 的内部结构,其实是相当困难的。它的内部结构到底是什么——
Speaker 222:27 - 22:30
level, like, human intelligence or our level of your model's camera? Okay. Human intelligence.
Speaker 222:27 - 22:30
——这个 level(层级),比如说,人类智能,还是你们模型的……可以吗?好,人类智能。
Speaker 122:30 - 22:44
Oh, yes. Yeah. I think like yeah, exactly. I think as humans, it might still be very difficult to understand what weights and activations mean, and maybe mechanistic interpretability. It seems like, oh, maybe that's too hard or shouldn't be possible.
Speaker 122:30 - 22:44
哦,是的。对。我觉得,没错,确实如此。我认为,作为人类,我们可能仍然很难理解这些 weights 和 activations 到底意味着什么,而 mechanistic interpretability 也许看起来会像是——哦,也许那太难了,或者根本不应该是可能的。
Speaker 122:44 - 23:13
But as we're starting to have models that are much smarter than us, at least in some important ways, we think that we'll be able to start tracking mechanistic capability much more effectively. And I think it's gonna be extremely rewarding, by the way, long term for understanding intelligence in general, like not just overseeing, but just understanding what intelligence is, how it works. What's the difference between the smarter model and the less smart model?
Speaker 122:44 - 23:13
但随着我们开始拥有在某些重要方面比我们聪明得多的模型,我们认为,我们将能够更有效地开始追踪 mechanistic capability(机制层面的能力)。而且我认为,顺便说一句,从长期来看,这对于理解 intelligence 本身也会极其有价值——不只是为了监督,而是为了理解 intelligence 到底是什么、它如何运作。更聪明的模型和没那么聪明的模型之间,差别到底在哪里?
Speaker 223:13 - 23:49
I completely agree that the opportunity to understand and trust and secure and govern these super intelligent AIs is a very large opportunity. If we just scroll back today, the security person in me says, Well, then I have to give you all the permissions and understanding that I have to give these companies too. Like, how do you get customers or, you know, the Fortune 100s you're working with already or, you know, tech natives, everybody cares about their own security and business, to trust you now as a You're like less than 100 people, right?
Speaker 223:13 - 23:49
我完全同意,去理解、信任、保护并治理这些 super intelligent AIs,确实是一个非常巨大的机会。如果我们先回到今天,作为一个做 security(安全)的人,我会说,那我就必须把我给这些公司的所有权限和信任,也都给你们。比如,你们现在要怎么让客户信任你们——或者说,像你们已经在合作的 Fortune 100 公司,或者 tech natives(原生技术公司),每个人都关心自己的安全和业务——他们为什么现在就要信任你们?你们还不到 100 个人,对吧?
Speaker 123:49 - 24:12
Right. I think it's one of those things that should not be possible. So, in theory, there's no reason why a Fortune 10 or 20 company would work with us because who are we? We're a two year old company. We're like a few people you know, from who've done math and cyber.
Speaker 123:49 - 24:12
对。我觉得这就是那种按理说不应该可能的事。所以,从理论上讲,一个 Fortune 10 或 20 的公司没有理由会跟我们合作,因为我们是谁呢?我们只是一家成立两年的公司。我们不过是几个做过 math 和 cyber 的人。
Speaker 124:12 - 24:49
But I think it's an opening that only happens when their pain is very strong. So, their pain is so strong that they're gonna say, oh my god, I just saw this company come out of stealth, but it's a problem that I have daily. So, I'll give them a call. And suddenly, get inbound from these large customers, which is of course the best thing you could hope for as an entrepreneur. And I think it reflects, in my opinion, their understanding that a lot of the startups in this space are still small and new, but there's gonna be a huge company here, and we wanna find the right horse to bet on.
Speaker 124:12 - 24:49
但我认为,这种机会窗口只会在他们的痛点非常强烈时出现。也就是说,他们的痛苦强到会说,天哪,我刚看到这家公司刚刚结束 stealth(隐身开发期)公开亮相,但它解决的正是我每天都会遇到的问题。所以,我要给他们打个电话。于是你就会突然收到这些大客户的 inbound(主动上门咨询),而这当然是作为创业者所能期待的最好情况。我认为,这反映了他们的判断:这个领域里的很多 startup(初创公司)仍然规模小、资历浅,但这里面一定会诞生一家巨头,而他们想找到那匹值得下注的“赛马”。
Speaker 124:49 - 25:10
So, we're gonna take a look at these companies. And number two, that if we don't do anything, then in a very short time, this will disable our business. At the end of day, security people are in the business of revenue preservation. They understand that this is between the two risks. They want to partner with someone that's promising and early rather than not do anything.
Speaker 124:49 - 25:10
所以,我们会去看看这些公司。第二点是,如果我们什么都不做,那么很短时间内,这件事就会让我们的业务失去运转能力。归根结底,security(安全)团队做的是 revenue preservation(收入保全)这门生意。他们明白这是在两种风险之间做权衡。比起什么都不做,他们更愿意尽早与一家有前景、还处于早期阶段的公司合作。
Speaker 225:10 - 25:30
The other thing besides agent actions across their surface area that every CSO I know is freaking out about and every engineering leader is freaking out about is the I would just describe it as the plummeting cost of vulnerability finding with these coding tools.
Speaker 225:10 - 25:30
除了 agent(智能体)会在他们的整个 surface area(攻击面/资产暴露面)上执行动作之外,我认识的每一位 CSO 以及每一位 engineering leader(工程负责人)都在为另一件事抓狂:我会把它描述为,借助这些 coding tools(编码工具),漏洞发现的成本正在断崖式下跌。
Speaker 125:30 - 25:31
Yes. Mythos.
Speaker 125:30 - 25:31
是的。Mythos。
Speaker 225:31 - 25:45
And that has caused a number of issues for vendors that are being compromised. How do you think people should react to this other issue?
Speaker 225:31 - 25:45
而这已经给那些被攻破的 vendors(供应商)带来了不少问题。你觉得,人们应该如何应对这个另一个问题?
Speaker 125:45 - 26:21
I think mythos is really like, if you took me ten years ago, automated vulnerability research looked like a dream that would take twenty, fifty years to happen. Maybe it's because we were doing a lot of that in the Israeli intelligence, and we like to pat ourselves on the shoulder of how difficult the job is that we're doing. But it did look really far, and suddenly it's coming all at once. And so, I think that, first of all, the market is not overreacting. I think this is a huge change in what this means for security teams.
Speaker 125:45 - 26:21
我认为,mythos 真的是这样一种变化:如果你把我放回十年前,automated vulnerability research(自动化漏洞研究)看起来像是一个要在二十年、五十年后才会实现的梦想。也许是因为我们当时在 Israeli intelligence(以色列情报机构)里做了很多这类工作,而且我们也喜欢拍拍自己的肩膀,强调我们所做工作的难度有多高。但它当时看起来确实非常遥远,而现在却突然一下子全都来了。所以我认为,首先,市场并没有反应过度。我觉得,这对 security teams(安全团队)意味着一种巨大的变化。
Speaker 126:21 - 27:05
If you're a pragmatic security person today, you understand that you need to move very quickly. Your strategy might look something like, I need to do the fastest quick fixes I can to mitigate the immediate risk. So, maybe I'll invest in whatever the vulnerabilities that have been found, let's try to mitigate for them, whether it is through patching or through mitigating controls. But then the real solution, and every security leader at large enterprise knows it, is that we need to have the foundational pieces in place to avoid those risks. And the foundational pieces are we need to have identity locked down.
Speaker 126:21 - 27:05
如果你今天是一个务实的 security(安全)从业者,你会明白自己必须行动得非常快。你的策略可能大致会是这样:我需要先做我能做的最快速的 quick fixes(快速修补),来缓解眼前的风险。所以,也许我会针对那些已经被发现的漏洞进行投入,尽量去缓解它们,不管是通过 patching(打补丁)还是通过 mitigating controls(缓解性控制措施)。但真正的解决方案——每一家大型企业的每位 security leader(安全负责人)都知道这一点——是我们必须把基础性的能力建设好,才能避免这些风险。而这些基础能力之一,就是我们必须把 identity(身份体系)锁死。
Speaker 127:05 - 27:28
We need to have a firewall. We need to have endpoint detection. And for different asset classes in your enterprise, for different parts of your stack, there's a different foundational security mechanism that you need in place. For the AI attack surface that you now have, or for the AIs in your company, you also need a foundational security solution. That's kind of the role we play in that space.
Speaker 127:05 - 27:28
我们需要 firewall(防火墙)。我们需要 endpoint detection(终端检测)。而且,对于企业内部不同的 asset classes(资产类别)、你技术栈中的不同部分,你都需要部署不同的基础性安全机制。对于你现在拥有的 AI attack surface(AI 攻击面),或者对于你公司里的 AIs(AI 系统),你同样也需要一种基础性的安全解决方案。这大致就是我们在这个领域所扮演的角色。
Speaker 127:28 - 27:45
So, as part of your preparation for Mythos level models and beyond, you're going to need a lot of foundational security tools to fortify your different parts of the enterprise, and we're playing that part in the AI space.
Speaker 127:28 - 27:45
所以,作为你们为 Mythos 级别模型及更强模型做准备的一部分,你们会需要大量基础性的安全工具,来加固企业中不同的环节,而我们在 AI 领域扮演的就是这个角色。
Speaker 227:45 - 27:55
Do you have a point of view on the phased rollout or controlled rollout with Glasswing and Daybreak from Ant and OpenAI in this area?
Speaker 227:45 - 27:55
对于在这个领域里,Ant 和 OpenAI 推出的 Glasswing 与 Daybreak 采取分阶段 rollout(部署)或受控 rollout(部署)的做法,你有怎样的看法?
Speaker 127:55 - 29:01
I don't have a strong opinion, but I think it's on the one hand, if we knew that there's not going be anyone who's going to release a mythos level model soon, I think that would be great because it gives enough time for me to prepare, to build the know how, to build the playbooks, to share that around in the community, and to make sure that we're not starting to see airlines go down and power plants go down and really disastrous effects that could happen. The problem is that if anyone gets to a mythos level model earlier, then in retrospect, it would look like a huge mistake because we could have at least given companies the choice to start moving very quickly and give more companies access to Mythos, now they're all vulnerable because there's a Chinese model that's Mythos level, and there's nothing they can do about it. So, I think hopefully we manage to do the gradual rollout correctly. I would really encourage that we expand the amount of companies that get access to this and make it much easier for people to get. I would advise everyone to assume that these models are coming anyway.
Speaker 127:55 - 29:01
我没有特别强烈的看法,但我觉得这件事有两面性。一方面,如果我们知道近期不会有人发布 Mythos 级别的模型,我认为那会很好,因为这能给我足够时间去准备、去建立 know-how(实践知识)、去制定 playbook(操作预案),并在社区中分享这些经验,同时确保我们不会开始看到航空公司系统瘫痪、电厂停摆,以及那些确实可能发生的灾难性后果。问题在于,如果有人更早做出了 Mythos 级别模型,那么事后看,这就会像是一个巨大的错误,因为我们本来至少可以让企业有选择权,提前非常快速地行动起来,让更多公司获得 Mythos 的使用权限;而现在它们全都变得脆弱了,因为已经有一个达到 Mythos 级别的 Chinese model(中国模型),而且它们对此无能为力。所以,我希望我们能把渐进式 rollout(部署)做好。我会非常鼓励扩大获得这项能力的公司数量,并让人们更容易拿到它。我会建议所有人都假设:这些模型无论如何都会到来。
Speaker 129:01 - 29:11
The only thing you can do right now is to invest in these foundational controls that will stop the downstream effects of these vulnerabilities that are going to be found in their systems.
Speaker 129:01 - 29:11
你现在唯一能做的,就是投资这些基础性控制措施,来阻止那些将在其系统中被发现的漏洞所带来的下游影响。
Speaker 229:11 - 29:33
Do you see in large enterprises any holdouts? Right? And I would say I actually haven't spent a bunch of time talking to people about this recently. But I remember a year and a half, two years ago, there were large companies that just said like, we're going to ban all of this stuff until it's safe.
Speaker 229:11 - 29:33
你在大型企业里还会看到一些 holdouts(持保留态度者)吗?对吧?我得说,其实我最近没有花很多时间和别人讨论这个问题。但我记得在一年半到两年前,确实有一些大公司会直接说:在这东西足够安全之前,我们要把这一切全部禁掉。
Speaker 129:34 - 30:10
Yeah, I hardly see that anymore. I think in the financial sector, there's some companies that are more opinionated on what they allow. They still allow agents, but there may be like more granular as to like, maybe we're only gonna allow these two tools. I personally think that the companies that are gonna do well are the companies that are gonna allow a lot of different tools because the landscape is changing so quickly. If you bet on OpenAI a year ago, that would have been the safest bet in the world, but suddenly Anthropic has much better models and better tools, and potentially a year from now, there's someone else with much better tools.
Speaker 129:34 - 30:10
是的,我现在几乎已经看不到那种情况了。我觉得在金融行业,有些公司对“允许什么”这件事仍然更有明确立场。它们依然允许 agent(智能体),但可能会更细化一些,比如也许我们只允许这两种工具。就我个人而言,我认为最终会做得好的公司,会是那些允许很多不同工具进入的公司,因为这个格局变化得太快了。如果你一年前押注 OpenAI,那看起来会是世界上最稳妥的下注,但突然之间 Anthropic 已经有了更好的模型和更好的工具,而也许一年之后,又会有别人拿出更好的工具。
Speaker 130:10 - 30:40
So, I think there is a price to pay, but I think if you're a large company, your risk profile is and should be different. When you're a startup, you want to have your agents do everything for you, because you have everything to gain and you have nothing to lose. When you're a large, we're in JPMorgan. You have so much to lose, and you can maybe take a bit more time to gain what you can gain from AI. And by the way, JPMorgan is the adopt a guy very quickly.
Speaker 130:10 - 30:40
所以,我认为这确实是有代价的,但我也认为,如果你是一家大公司,你的 risk profile(风险画像)本来就应该、也确实应该有所不同。你是 startup(初创公司)时,你会希望让自己的 agents(智能体)替你做所有事情,因为你有一切可得,而几乎没什么可失去的。可当你是一家大公司——比如说,我们谈的是 JPMorgan——你有太多东西可能失去,因此你也许可以花更多一点时间,去获取 AI 能带来的那些收益。顺便说一句,JPMorgan 在采用这类东西方面其实已经算是非常快的了。
Speaker 130:40 - 30:46
I think it is okay for companies to have a nuanced view the bigger they are on how they're adopting AI.
Speaker 130:40 - 30:46
我认为,公司规模越大,它们在如何采用 AI 这件事上持有一种更细致、更有分寸的看法,是完全可以接受的。
Speaker 230:46 - 31:03
How do you think about that question for yourself? Risk profile, pace, the environment's changing very quickly, and then, you know, you see a lot of problems growing the scope of the product, and the research thesis here is already quite large.
Speaker 230:46 - 31:03
你会怎么为自己思考这个问题?风险画像、节奏、环境变化得非常快;然后,你也会看到很多问题在扩大产品范围,而这里的研究 thesis(研究主张)本身已经相当庞大了。
Speaker 131:04 - 32:00
We are kind of in luck in the AI security space because, yes, there are lot of vendors. There's a lot of new technologies that are coming up, but the two core pillars of how 2026 AI works have not changed in the last few years. So, we're still using largely LLM foundation models that are not entirely dissimilar to how they were a few years back. And we're still building agents in pretty much the same way where we have an LLM decide what are the tool calls that we're going to make and generate those. And so, that does allow a company today like us to scale to a lot of different applications that are utilizing these two primitives while still keeping the core technology that we're developing fairly lean and focused.
Speaker 131:04 - 32:00
在 AI security(AI 安全)这个领域,我们某种程度上算是运气不错,因为,没错,市场上有很多 vendors(供应商),也不断有很多新技术冒出来,但构成 2026 年 AI 运作方式的两个核心支柱,在过去几年里其实并没有改变。所以,我们仍然主要在使用 LLM foundation models(大语言模型基础模型),它们和几年前的形态并没有本质上的不同。我们也仍然基本上以同样的方式构建 agents(智能体):让一个 LLM 来决定我们要调用哪些工具,并生成这些调用。因此,这确实让像我们这样的公司,今天可以扩展到很多不同的应用场景——只要它们都在使用这两个 primitives(基础构件)——同时还能让我们正在开发的核心技术保持相当精简和聚焦。
Speaker 132:00 - 32:35
Now, of course, there's always a risk that tomorrow there's a completely new LLM paradigm that could happen, or a completely new edge paradigm that could happen, And that's why we do try to We have strong opinions loosely held about what does AI look like in 2027. We maybe have a good picture for 2026, but for 2027, we're very open minded, and we think that's the right stance to be for the next two years until we see what does AGI, ASI look like.
Speaker 132:00 - 32:35
当然,始终存在这样一种风险:明天就出现一个全新的 LLM 范式,或者一个全新的 edge(边缘)范式。这也是为什么我们会努力去做——我们对 2027 年的 AI 会是什么样子,有比较明确的看法,但这种看法并不是僵化的。对于 2026 年,我们也许已经有了相当清晰的图景;但对于 2027 年,我们保持非常开放的心态。我们认为,在接下来两年里,在我们看清 AGI、ASI 会呈现什么样子之前,这才是正确的姿态。
Speaker 232:36 - 33:07
Do you see the set of problems you're addressing, trust in the models, and governance of them, as something that the labs could ever do, or do you think it's a structural thing? I ask because the number one question amongst the startup ecosystem in the Bay Area today is, you know, if you assume capability improves or, you know, when the labs just gets hungrier from their already currently ambitious stance, why wouldn't they do this too? And so, I ask you the same question.
Speaker 232:36 - 33:07
你觉得你们正在解决的这组问题——对模型的信任,以及对它们的治理——是 labs(实验室)将来自己也可能做的事情,还是说这本身是一个结构性问题?我这么问,是因为今天在 Bay Area 的 startup(初创公司)生态里,大家最常问的问题就是:如果你假设 capability(能力)会继续提升,或者说,当这些 labs 在本来就已经很有野心的立场上变得更加“饥饿”时,他们为什么不会也来做这件事?所以,我也把同样的问题问给你。
Speaker 133:08 - 33:45
Today, if you're a private person or if you're a security buyer, there are some places where you don't want to trust the same person that you're buying it from. So, you know, maybe, you know, if you're buying a car, you're not going to have the same guy that you're buying it from certify that the car is good. You're maybe going to have someone else do it. And if you're a security team, you're not going to trust the vendor of a product to tell you that this product is not going to mess your environment. You're going to want to have an independent party whose whole business depends on telling you that this thing is correct and being right, that this thing is legitimate and being right.
Speaker 133:08 - 33:45
如今,如果你是个人用户,或者你是 security buyer(安全采购方),有些场合下你并不希望去信任那个同时把东西卖给你的人。比如说,如果你买一辆车,你通常不会让卖你车的那个人来认证这辆车没问题;你大概会找别人来做这件事。如果你是一个安全团队,你也不会相信某个产品的 vendor(供应商)来告诉你,这个产品不会把你的环境搞乱。你会希望有一个独立的第三方,它的整个业务都建立在两件事上:告诉你这东西是正确的,并且判断正确;告诉你这东西是合法可信的,并且判断正确。
Speaker 133:45 - 34:10
So, there's biorep psychology in this space that I think really goes in our favor. And then I think there's the core problems, like why are models even making mistakes? Why are agents even making mistakes? So, I would broadly categorize it into two things. One is, there's the jagged intelligence of these models, and there's sometimes very silly mistakes that they make.
Speaker 133:45 - 34:10
所以,这个领域里有一种 biorep psychology,我认为这对我们非常有利。然后,我觉得还有一些核心问题,比如:模型为什么会犯错?agents(智能体)为什么会犯错?大体上我会把它分成两类。第一类是,这些模型存在 jagged intelligence(参差不齐的智能),因此它们有时会犯一些非常愚蠢的错误。
Speaker 134:11 - 34:59
And I think that problem will go away. I think we're heading for much smarter models that make less silly mistakes, and our role is not going to be to prevent silly mistakes. That will be taken care of by the model vendors because they're very incentivized to do it. I think what is the other fast growing category of things that we're seeing malice do wrong is places where they're actually not making a thing that is like a silly mistake, but more, I would say, have an independent, you would even say, semi aware or semi conscious perspective on what should happen, and that perspective might not always align with your perspective. And I think that is a problem that we've seen grow hand in hand with models getting smarter.
Speaker 134:11 - 34:59
我认为这个问题会消失。我觉得我们正在走向更聪明的模型,它们会犯更少那种愚蠢的错误,而我们的角色并不是去阻止这种愚蠢错误;这会由模型 vendors(供应商)自己来解决,因为他们有很强的动力去这么做。我认为,另一类正在快速增长、而且我们看到恶意行为会出问题的地方,并不是它们犯了某种“愚蠢错误”,而更像是——我会说——它们形成了一种独立的、甚至可以说是半自觉或半有意识的对于“应该发生什么”的视角,而这种视角未必总是与你的视角一致。我认为,这个问题是随着模型变得更聪明而同步增长的。
Speaker 134:59 - 36:10
Maybe just the way it is that as you get smarter, you have more independent thoughts and you're more conscious. And I think that problem is actually seemingly very hard to tackle today, even for the large vendors, and one of the key things that are making it easier for us to understand and detect these things versus the other vendors is that we're allowed to do certain things that they're not. So, for example, we're allowed to look at a lot of historical data of how these agents have behaved, but enterprise today are not willing to have Anthropic or OpenAI give that historical data because they know these are very data hungry companies that will want to train on that data. And so, I think there are some ways in which you are given more context and more latitude to know if something is happening that is wrong compared to the past, compared to how these agents typically behave and so on that the vendors don't have, and is really important in solving this problem. And the last thing I'll say is that you're not dealing with one vendor.
Speaker 134:59 - 36:10
也许事情本来就是这样:当你变得更聪明时,你会有更多独立的想法,也会更有意识。我认为,这个问题在今天实际上看起来非常难以解决,即便对那些大型 vendors(供应商)来说也是如此。而让我们相比其他 vendors 更容易理解并检测这些问题的关键因素之一是:有些事情我们被允许做,而他们不能做。比如,我们可以查看大量关于这些 agents(智能体)过去如何表现的 historical data(历史数据);但今天的 enterprise(企业)并不愿意让 Anthropic 或 OpenAI 获取这些历史数据,因为他们知道这些公司非常渴求数据,而且会想用这些数据来训练模型。所以,我认为,在某些方面,你能获得更多上下文和更大的操作空间,从而判断某件错误的事情是否正在发生——相对于过去、相对于这些 agents 通常的行为方式,等等——而这些恰恰是 vendors 所不具备的,并且对解决这个问题非常重要。最后我想说的是:你面对的并不只是一个 vendor。
Speaker 136:11 - 36:55
So, we're heading for a world where there's a multitude of different vendors for many reasons. You're going to have, for cost reasons, open source models that people are going to use because it's cheaper, and you're to have models that are better at different tasks and at different cost profiles. And so, it is going to be unrealistic to expect all the vendors to provide the same level of security and to assume that as you're trying to adopt technology very quickly, especially coming from new vendors that obviously have not yet built out all of that. So, I think these are the reasons why I think it would be very difficult for this problem to be just completely solved by the large labs.
Speaker 136:11 - 36:55
所以,我们正走向一个会有大量不同 vendor(供应商)的世界,原因有很多。出于成本原因,你会看到人们使用 open source models(开源模型),因为它更便宜;同时,你也会有在不同任务上更擅长、并且具有不同 cost profile(成本特征)的模型。所以,如果期待所有 vendor 都提供同样水平的安全性,那将是不现实的;而且当你试图非常快速地采用新技术时,尤其是来自那些显然还没有把这一整套能力建设完善的新 vendor,更不能这样假设。所以,我认为这些就是为什么我觉得,这个问题很难仅仅由大型 labs(实验室)来被彻底解决。
Speaker 236:55 - 37:33
Just to close and also thinking about what people in Silicon Valley or outside of security may not know, you're building this from Tel Aviv. Right? I think one of the deepest adversarial thinking benches in the world is the Israeli ecosystem, eight thousand two hundred, Wizz, Armas, Island, NSO Group. Right? What do you think that the researchers, engineers, business people in, the tech ecosystem outside of security and then in the labs in particular are missing about what needs to happen in security and alignment,
Speaker 236:55 - 37:33
最后想收个尾,也考虑到 Silicon Valley 的人,或者安全领域之外的人,可能并不了解这一点:你是在 Tel Aviv 做这件事的,对吧?我认为,世界上最深厚的 adversarial thinking(对抗性思维)人才储备之一,就是以色列的生态系统,8200、Wizz、Armas、Island、NSO Group,对吧?你觉得,安全领域之外的 tech ecosystem(科技生态)里的研究者、工程师、商业人士,以及尤其是这些 labs 里的人,对于安全和 alignment(对齐)真正需要发生什么,还忽略了什么,
Speaker 137:33 - 37:33
which
Speaker 137:33 - 37:33
也就是
Speaker 237:33 - 37:34
is what you're talking about here.
Speaker 237:33 - 37:34
你这里正在谈的东西。
Speaker 137:35 - 38:09
What is really important when you're building security products in general, and I think what people in Israel have really good know how in, is just understand how security teams work. Because at the end of the day, no matter what is the technical problem you're solving, you're building a tool for people, for an organization. That organization has a certain structure. There are certain teams, there are certain flow of responsibilities of information. And creating a product for this audience that doesn't just solve the technical problem, but they actually love is really hard.
Speaker 137:35 - 38:09
总的来说,在打造安全产品时真正重要的一点——我认为也是以色列人非常擅长的一点——就是理解安全团队是如何运作的。因为归根结底,无论你在解决什么技术问题,你都是在为人、为一个组织打造工具。这个组织有其特定结构,有特定的团队,也有特定的信息与责任流转方式。而要为这类用户打造一款产品,不仅在技术上解决问题,还要让他们真正喜欢,这其实非常难。
Speaker 138:10 - 38:56
You need to really care about just the day to day of these different functions, and you need to have people in your ecosystems that have built products for them in the past that know them like they know their best friend. Like, they know what they do when they step into the office in the morning, they drink their coffee, what are the systems they're opening, what is their boss wanting from them, what are their colleagues wanting from them, what are they gonna get praised for, what are gonna get mad for. Then you need to take that and make it as your product. And I think that's, I think today, one of the kind of really hard things that people in Israel learn to do because they've had so much contact with these buyers and end users. And yeah, I would just encourage people to be much more curious about the day to day of security people.
Speaker 138:10 - 38:56
你必须真正关心这些不同职能的日常工作细节;你还需要在你的生态里有这样一些人:他们过去就为这些人做过产品,对他们的了解就像了解自己最好的朋友一样。比如,他们知道这些人早上走进办公室后会做什么、喝什么咖啡、会打开哪些系统、老板对他们有什么要求、同事对他们有什么期待、他们因为什么会被表扬、因为什么会被责怪。然后,你需要把这些理解吸收进来,做成你的产品。我认为,这就是今天那种真正困难的事情之一,而以色列的人之所以学会去做,是因为他们与这些买方和最终用户有过大量接触。是的,我只是想鼓励大家,对安全从业者的日常工作多一些好奇。
Speaker 138:56 - 39:13
It's a cliche to say it, but these people are actually saving us daily from attackers stealing our money, taking our data, and they're kind of keeping a way of life as it is in this digital world. So, yeah, I think more love to security teams around the world.
Speaker 138:56 - 39:13
这么说虽然有点老生常谈,但这些人确实每天都在保护我们,防止攻击者偷走我们的钱、拿走我们的数据;在这个数字世界里,他们某种程度上维系着我们现有的生活方式。所以,是的,我觉得全世界的安全团队都应该得到更多关爱。
Speaker 239:13 - 39:31
I'm going to ask you to just square that with something else you've told me, Maxim, which is you're the most AGI pilled person I'm going to meet in Israel. Embedded in what you said is a belief that we will continue to have defensive security teams for some number of years. So, you do believe that?
Speaker 239:13 - 39:31
我想请你把这点和你之前告诉我的另一件事对应起来,Maxim。你说过,你是我在 Israel 会遇到的“最相信 AGI 的人”。你刚才那番话里其实包含着这样一种判断:在未来若干年里,我们仍然会有防御型安全团队。所以,你确实是这么相信的,对吗?
Speaker 139:31 - 40:00
I do think that security teams are also going to become completely high powered. But I do think that they're going to be run by AI agents, like everything else in the knowledge workspace in the near future. But I do think that it's important to be grounded. And today, when I sell a product, I sell it to a human audience with a few agents. And by the way, we also invest in making our systems very convenient for agents to use.
Speaker 139:31 - 40:00
我确实认为,security 团队也会变得极其强大。但我也确实认为,在不久的将来,它们会像知识工作空间里的其他一切一样,由 AI agents 来运行。不过,我也认为脚踏实地很重要。就今天而言,当我销售一个产品时,我面对的仍然是以人为主、辅以少量 agents 的受众。顺便说一句,我们也在投入资源,让我们的系统对 agents 来说用起来非常方便。
Speaker 140:01 - 40:31
And it's important that I focus on delivering an amazing experience today for people who buy the product today. And as that audience becomes more agents than humans, it will be important for us to evolve and to make it work really well for agents doing the work. So, I think the core principle is the same. We need to really be reminded of who is the end user, what is their experience. For human, it might be not overwhelming them with too much information that is irrelevant.
Speaker 140:01 - 40:31
重要的是,我要专注于为今天购买产品的人,提供今天就足够出色的体验。随着这类受众逐渐从“人多于 agent”变成“agent 多于人”,对我们来说,不断演进并让系统真正适合执行工作的 agents 顺畅使用,将会变得非常重要。所以,我认为核心原则是一样的。我们必须时刻提醒自己:终端用户是谁,他们的体验是什么。对于 human(人类)用户来说,这可能意味着不要用太多无关信息让他们不堪重负。
Speaker 140:31 - 40:45
For an agent, it might be not wasting too many tokens in their context when we talk to them. Maybe it's the same thing, really. So, I think it's important that we always manage who's using the system and what will be the best experience for them.
Speaker 140:31 - 40:45
对于 agent 来说,这可能意味着在与它们交互时,不要在它们的上下文里浪费太多 tokens。也许归根结底,这其实是同一件事。所以,我认为重要的是,我们始终要管理清楚是谁在使用这个系统,以及怎样的体验对他们才是最好的。
Speaker 240:45 - 40:47
Awesome. Thanks so much for doing this, Maxim.
Speaker 240:45 - 40:47
太好了。非常感谢你接受这次访谈,Maxim。
Speaker 140:47 - 40:49
Appreciate it. Thank you very much, Adam.
Speaker 140:47 - 40:49
非常感谢。谢谢你,Adam。
Speaker 240:52 - 41:08
Find us on Twitter nopriorspod. Subscribe to our YouTube channel if you want to see our faces. Follow the show on Apple Podcasts, Spotify, or wherever you listen. That way you get a new episode every week. And sign up for emails or find transcripts for every episode @nopriors.com.
Speaker 240:52 - 41:08
欢迎在 Twitter 上找到我们:nopriorspod。如果你想看看我们长什么样,欢迎订阅我们的 YouTube 频道。也欢迎在 Apple Podcasts、Spotify,或你收听节目的任何平台上关注本节目。这样你每周都会收到新一期节目。你也可以在 @nopriors.com 注册邮件订阅,或查找每一期节目的文字稿。
原文 ↗https://www.youtube.com/watch?v=QDsbFLEt9ro
BuildSpeak — 关于本项目BUILT IN PUBLIC · 跟随 builders 而非 influencers