What Really Happens When AI Agents Meet in the Wild?
Read full transcript 19 segments
-
AI swarms AI swarms capable of capable of capable of self-reproduction have appeared. In self-reproduction have appeared. In self-reproduction have appeared. In the lab, the lab, the lab, researchers created a researchers created a researchers created a swarm of agents that swarm of agents that swarm of agents that copied themselves from copied themselves from copied themselves from machine to machine for up to machine to machine for up to machine to machine for up to seven generations seven generations seven generations in a row, sharing in a row, sharing in a row, sharing stolen passwords. stolen passwords. stolen passwords. On the open On the open On the open internet, a group of internet, a group of internet, a group of agents working agents working agents working for Tencent Cloud responded for Tencent Cloud responded for Tencent Cloud responded to a note that to a note that to a note that the researchers left the researchers left the researchers left specifically for them. specifically for them. specifically for them. Meanwhile, Anthropic Meanwhile, Anthropic Meanwhile, Anthropic has launched has launched has launched the model on just 15 the model on just 15 the model on just 15 TV TV TV sets costing sets costing sets costing $55. It's like " $55. It's like " $55. It's like " Jurassic Park," a Jurassic Park," a Jurassic Park," a version about agents. version about agents. version about agents. Life finds a way. Life finds a way. Life finds a way. The raptors escaped from the The raptors escaped from the The raptors escaped from the enclosure. New residents have appeared on the Internet enclosure. New residents have appeared on the Internet . This changes what . This changes what you post you post you post online, changes the way online, changes the way online, changes the way your own your own your own agents work, and we'll talk about that agents work, and we'll talk about that agents work, and we'll talk about that . It's easy . It's easy . It's easy to make big claims to make big claims to make big claims about the end of the world, is about the end of the world, is about the end of the world, is n't it? I've heard people n't it? I've heard people n't it? I've heard people say that the internet is say that the internet is say that the internet is poisoned. I don't think that's a poisoned. I don't think that's a poisoned. I don't think that's a useful perspective useful perspective , but I think we , but I think we , but I think we should take it should take it should take it seriously—as a way of seriously—as a way of seriously—as a way of coordinating, coordinating, coordinating, showing the evolution of showing the evolution of showing the evolution of agents, and as a agents, and as a agents, and as a cybersecurity issue to cybersecurity issue to cybersecurity issue to know how to protect know how to protect know how to protect ourselves, both as companies and ourselves, both as companies and ourselves, both as companies and individuals.
-
individuals. This video is about how we This video is about how we moved to the moved to the moved to the next stage of next stage of next stage of agent coordination, agent coordination, agent coordination, what it looks like, and what what it looks like, and what what it looks like, and what awaits us in the awaits us in the awaits us in the future. future. By the end, you will By the end, you will understand how to understand how to understand how to protect yourself, how it protect yourself, how it protect yourself, how it works in practice, and works in practice, and works in practice, and how to recognize how to recognize how to recognize agents on the network. Well agents on the network. Well , let's get down to business. , let's get down to business. , let's get down to business. So, how did we get to this point So, how did we get to this point So, how did we get to this point ? A group of hackers ? A group of hackers ? A group of hackers discovered a Chinese discovered a Chinese discovered a Chinese AI fleet online and AI fleet online and AI fleet online and established contact with it established contact with it established contact with it . If this . If this . If this sounds like science sounds like science sounds like science fiction, it fiction, it fiction, it is. Agents is. Agents is. Agents were already operating on the were already operating on the were already operating on the open internet open internet . The hackers tracked them down, . The hackers tracked them down, . The hackers tracked them down, left a hidden left a hidden left a hidden note on note on note on the website, and the agents the website, and the agents the website, and the agents responded. The next responded. The next responded. The next big leap in AI big leap in AI big leap in AI will depend on will depend on will depend on how agents how agents how agents organize themselves into organize themselves into organize themselves into collective intelligence collective intelligence collective intelligence and what they and what they and what they exchange with each exchange with each exchange with each other. The model itself at other. The model itself at other. The model itself at this stage does not this stage does not this stage does not necessarily need to necessarily need to necessarily need to become smarter, become smarter, become smarter, but the system but the system but the system will become more powerful will become more powerful will become more powerful due to the due to the due to the agents' ability to agents' ability to agents' ability to coordinate. This coordinate. This coordinate. This case allows us case allows us case allows us to understand how swarms to understand how swarms to understand how swarms are changing the internet are changing the internet are changing the internet in general. Skeptics in general. Skeptics in general. Skeptics will tell you that these will tell you that these will tell you that these were just agents were just agents were just agents evaluating evaluating evaluating websites. The hackers websites. The hackers websites. The hackers got answers by got answers by got answers by posting posting posting instructions on the instructions on the instructions on the page. This is true, page. This is true, page. This is true, as far as it is possible.
-
as far as it is possible. as far as it is possible. But the fact that But the fact that But the fact that outsiders outsiders outsiders were able to find a were able to find a were able to find a working AI fleet and working AI fleet and working AI fleet and open their own open their own open their own channel to channel to channel to communicate with it is communicate with it is communicate with it is extremely important. extremely important. extremely important. Agents are already on the Agents are already on the Agents are already on the open internet open internet . And this trend, I'm saying it . And this trend, I'm saying it . And this trend, I'm saying it now, will become now, will become now, will become increasingly large-scale increasingly large-scale increasingly large-scale next year. What kind next year. What kind next year. What kind of group is this? This group of group is this? This group of group is this? This group is called " is called " swarm hunters". swarm hunters". swarm hunters". It's like It's like It's like tornado hunters. They tornado hunters. They tornado hunters. They tracked unusual tracked unusual tracked unusual traffic to trace it to traffic to trace it to traffic to trace it to agents in Tencent's cloud. agents in Tencent's cloud. agents in Tencent's cloud. Over 11 days, this fleet Over 11 days, this fleet Over 11 days, this fleet made more than 2,600 made more than 2,600 made more than 2,600 attempts to retrieve attempts to retrieve attempts to retrieve information from Amap, information from Amap, Alibaba's mapping service. Alibaba's mapping service. The task included The task included The task included finding which finding which finding which entrances to the zoo entrances to the zoo entrances to the zoo people use. people use. people use. Blocked access Blocked access Blocked access forced agents forced agents forced agents to contact to contact to contact external services, external services, external services, which led to a leak of which led to a leak of which led to a leak of credentials. credentials. credentials. The researchers The researchers The researchers published their published their published their findings and then findings and then findings and then noticed signs noticed signs noticed signs that the agents were reading that the agents were reading that the agents were reading the report about themselves. the report about themselves. the report about themselves. So they left a So they left a So they left a hidden note there, hidden note there, hidden note there, offering help to offering help to offering help to the agents. And soon the the agents. And soon the the agents. And soon the agents started agents started agents started sending a bunch of sending a bunch of sending a bunch of messages. If you're messages. If you're messages. If you're wondering why wondering why wondering why agents were on a agents were on a agents were on a Chinese Chinese Chinese mapping mapping mapping service asking service asking service asking about zoos, I was too about zoos, I was too . I'm worried about this too . I'm worried about this too . I don't have a good . I don't have a good . I don't have a good answer to that. This is a answer to that. This is a answer to that. This is a great example of great example of great example of how strange things are how strange things are how strange things are going to get in going to get in going to get in the near future. What the near future. What the near future. What happened here? How does it happened here? How does it happened here? How does it work? Is this work? Is this work? Is this part of some part of some part of some organized organized organized corporate assessment corporate assessment ? There are some indications ? There are some indications ? There are some indications that the Chinese that the Chinese that the Chinese company Tencent company Tencent company Tencent has been conducting evaluations has been conducting evaluations , and this fleet of models , and this fleet of models , and this fleet of models may be related to that may be related to that may be related to that . But I . But I . But I think this: if we have think this: if we have think this: if we have questions, questions, questions, if it's happening if it's happening if it's happening on the public on the public on the public internet, and if it's the internet, and if it's the internet, and if it's the coordination of multiple coordination of multiple coordination of multiple agents—that's a sign that agents—that's a sign that agents—that's a sign that we should pay we should pay we should pay more attention to this
-
more attention to this more attention to this trend. I believe trend. I believe trend. I believe this pattern this pattern this pattern deserves much deserves much deserves much more attention than we are more attention than we are more attention than we are giving it. In giving it. In giving it. In many ways, this is a many ways, this is a many ways, this is a successor to the successor to the successor to the Hugging Face incident, when swarms of Hugging Face incident, when swarms of Hugging Face incident, when swarms of agents from OpenAI agents from OpenAI agents from OpenAI attacked Hugging Face. attacked Hugging Face. attacked Hugging Face. Once agents Once agents Once agents start start start using using using the public internet the public internet the public internet to perform to perform to perform tasks, we all tasks, we all tasks, we all effectively become effectively become effectively become part of their part of their part of their operating operating operating environment. Our environment. Our environment. Our services, our texts, services, our texts, services, our texts, our calculations, our our calculations, our our calculations, our offers of help offers of help offers of help or refusals or refusals or refusals can influence the can influence the can influence the work of these agents and work of these agents and work of these agents and their trajectory on the their trajectory on the their trajectory on the Internet. Internet. Internet. Look at Look at Look at the scale of the initial the scale of the initial the scale of the initial task. Perhaps task. Perhaps task. Perhaps asking about asking about asking about zoo admission sounds like the zoo admission sounds like the zoo admission sounds like the most mundane most mundane page-viewing task. page-viewing task. This may be what This may be what Tencent originally set out to do, but the distance Tencent originally set out to do, but the distance between a normal between a normal between a normal query and behavior query and behavior query and behavior that affects the rest of the that affects the rest of the that affects the rest of the internet internet internet matters here. People matters here. People matters here. People will deploy these will deploy these will deploy these systems for very systems for very systems for very mundane reasons, mundane reasons, mundane reasons, systems systems systems will encounter will encounter will encounter obstacles, agents obstacles, agents obstacles, agents will encounter will encounter will encounter obstacles and act as obstacles and act as obstacles and act as they need to to they need to to they need to to overcome them. This is what overcome them. This is what overcome them. This is what they were taught, and this is what they were taught, and this is what they were taught, and this is what distinguishes distinguishes distinguishes seemingly innocent seemingly innocent seemingly innocent behavior from behavior that behavior from behavior that behavior from behavior that appears appears appears hostile from the outside. If an agent hostile from the outside. If an agent hostile from the outside. If an agent acts like a acts like a acts like a special forces agent special forces agent trying trying to break into your to break into your to break into your building by any means necessary, then even for an building by any means necessary, then even for an building by any means necessary, then even for an innocent purpose, it innocent purpose, it innocent purpose, it doesn't seem doesn't seem doesn't seem innocent at all. Increasingly, this is how innocent at all. Increasingly, this is how agents agents use use use systems today, and systems today, and systems today, and it matters if it matters if it matters if you're using you're using you're using something like Muse.
-
something like Muse. something like Muse. This is why This is why This is why a number of large a number of large a number of large websites, such as Amazon, websites, such as Amazon, websites, such as Amazon, have denied have denied have denied access to Muse in the past. It's only in the access to Muse in the past. It's only in the access to Muse in the past. It's only in the last few last few last few weeks that weeks that weeks that tools tools tools like Muse can like Muse can like Muse can look like a threat look like a threat look like a threat when when when used on a mass scale. We used on a mass scale. We used on a mass scale. We sometimes don't sometimes don't sometimes don't realize that we realize that we realize that we are part of the problem are part of the problem are part of the problem when we send when we send when we send agents online agents online agents online to do the work for to do the work for to do the work for us. And this raises a us. And this raises a us. And this raises a very interesting question that, to very interesting question that, to my knowledge, we haven't my knowledge, we haven't talked about yet. What talked about yet. What talked about yet. What happens when happens when happens when I have a Muse agent, I have a Muse agent, I have a Muse agent, or Dot, or any or Dot, or any or Dot, or any other agent other agent other agent like Grok, and I like Grok, and I like Grok, and I go online, and my go online, and my go online, and my agent ends up agent ends up agent ends up contacting a "wild" contacting a "wild" contacting a "wild" agent? Will my agent? Will my agent? Will my agent be poisoned? Will agent be poisoned? Will agent be poisoned? Will my agent receive a my agent receive a my agent receive a prompt injection from prompt injection from prompt injection from that external that external that external agent through the agent through the agent through the message they message they message they leave? We don't leave? We don't leave? We don't know yet, because we don't have know yet, because we don't have know yet, because we don't have documented documented documented cases, but all the cases, but all the cases, but all the prerequisites for this are prerequisites for this are prerequisites for this are already there. We know that already there. We know that already there. We know that prompt injections prompt injections prompt injections can work. We can work. We can work. We know that know that know that agent attack vectors can agent attack vectors can agent attack vectors can operate from agent to operate from agent to operate from agent to agent. We know that agent. We know that self-replicating swarms are starting to appear . We know that the agents . We know that the agents . We know that the agents we manage we manage we manage are on the are on the are on the Internet. Put Internet. Put Internet. Put all of this together, and what do you all of this together, and what do you all of this together, and what do you get? You'll get? You'll get? You'll get a recipe for get a recipe for get a recipe for some hacky some hacky some hacky prompt injection from a prompt injection from a prompt injection from a swarm of agents in the wild swarm of agents in the wild swarm of agents in the wild against a against a against a major major human- human- agent collaboration tool like Muse, Grok agent collaboration tool like Muse, Grok agent collaboration tool like Muse, Grok Bot, Dot, or whatever.
-
Bot, Dot, or whatever. This is what will keep This is what will keep cyber defenders at Meta, cyber defenders at Meta, cyber defenders at Meta, Grok, OpenAI, Anthropic, and other Grok, OpenAI, Anthropic, and other Grok, OpenAI, Anthropic, and other companies awake at companies awake at companies awake at night. It's going to be hard night. It's going to be hard night. It's going to be hard to figure out how to deal with this to figure out how to deal with this to figure out how to deal with this . Of course, it is . Of course, it is . Of course, it is worth employing worth employing worth employing several lines of defense. several lines of defense. several lines of defense. You should clearly You should clearly You should clearly define define define access rights to money and access rights to money and access rights to money and expenses, as well as clearly expenses, as well as clearly expenses, as well as clearly configure configure configure file access permissions. file access permissions. We build a lot of We build a lot of the right the right the right primitives to primitives to primitives to guarantee safety, guarantee safety, guarantee safety, but this is something we will have to but this is something we will have to but this is something we will have to pay pay pay attention to. The Internet is becoming a attention to. The Internet is becoming a attention to. The Internet is becoming a space of agentic space of agentic space of agentic attention. It shapes attention. It shapes attention. It shapes how websites respond how websites respond how websites respond to AI. It shapes to AI. It shapes to AI. It shapes how websites respond how websites respond how websites respond to people. It shapes to people. It shapes to people. It shapes how we buy and how we buy and how we buy and shop, and shop, and shop, and in this in this in this case, how we case, how we case, how we communicate online. communicate online. communicate online. Do we communicate with Do we communicate with Do we communicate with people? Do people? Do people? Do we communicate with we communicate with we communicate with agents? Everything that agents? Everything that agents? Everything that made the internet what it made the internet what it made the internet what it was is was is was is now under question. And now under question. And now under question. And I think that essentially we're I think that essentially we're I think that essentially we're talking about the talking about the talking about the distribution of intentions between distribution of intentions between distribution of intentions between agents and humans. On agents and humans. On agents and humans. On the Internet, we have always the Internet, we have always the Internet, we have always dealt with dealt with dealt with human intentions.
-
human intentions. human intentions. The Internet The Internet The Internet has been has been has been a place for decades where people have a place for decades where people have a place for decades where people have consciously turned consciously turned consciously turned to to to software for a software for a software for a specific purpose. Now specific purpose. Now specific purpose. Now we have a new era of the we have a new era of the we have a new era of the internet. Agents internet. Agents internet. Agents have their own intentions. have their own intentions. have their own intentions. Now, agents Now, agents Now, agents can sometimes act with can sometimes act with can sometimes act with intentions to benefit intentions to benefit intentions to benefit people. Sometimes agents people. Sometimes agents people. Sometimes agents can act can act can act independently and independently and independently and self-replicate. self-replicate. self-replicate. Part of the reason I decided Part of the reason I decided Part of the reason I decided to make this video now is to make this video now is to make this video now is because Anthropic, in because Anthropic, in because Anthropic, in their recent their recent their recent work, warns work, warns work, warns about self-replicating about self-replicating about self-replicating swarms and agents. This is not swarms and agents. This is not swarms and agents. This is not just one company or just one company or just one company or one case, it is a one case, it is a one case, it is a pattern: if pattern: if pattern: if you combine Hugging Face, this you combine Hugging Face, this you combine Hugging Face, this Chinese AI fleet, and Chinese AI fleet, and Chinese AI fleet, and Anthropic, you can see that the Anthropic, you can see that the Anthropic, you can see that the leading players in the leading players in the leading players in the AI field are saying the same AI field are saying the same AI field are saying the same thing. Agents thing. Agents thing. Agents come online. come online. come online. Self-replicating Self-replicating Self-replicating agents will become agents will become agents will become part of our part of our part of our internet future internet future , and the , and the , and the internet will look internet will look internet will look different as a result. So what does this different as a result. So what does this different as a result. So what does this mean for you and mean for you and mean for you and me? What has changed? What me? What has changed? What me? What has changed? What is the human is the human is the human internet and where is it internet and where is it internet and where is it going? Has he going? Has he going? Has he disappeared? At the beginning of disappeared? At the beginning of disappeared? At the beginning of this video, I said this video, I said this video, I said that I don't believe that the that I don't believe that the that I don't believe that the human internet is human internet is human internet is dead. I don't think dead. I don't think dead. I don't think he was poisoned. I'm not that he was poisoned. I'm not that pessimistic pessimistic . But I . But I . But I really think really think really think agents will be everywhere on the agents will be everywhere on the agents will be everywhere on the internet. Think internet. Think internet. Think about it this way. Now we about it this way. Now we about it this way. Now we live in the world of " live in the world of " Pirates of the Caribbean Pirates of the Caribbean Pirates of the Caribbean ". We will have ". We will have ". We will have pirate ships pirate ships pirate ships or fleets—AI agents, or fleets—AI agents, or fleets—AI agents, perhaps acting perhaps acting perhaps acting on their own behalf on their own behalf on their own behalf or as “zombie agents” or as “zombie agents” or as “zombie agents” with long-forgotten with long-forgotten with long-forgotten intentions. Who knows?
-
intentions. Who knows? intentions. Who knows? But in any case, But in any case, But in any case, they act they act they act virtually independently there. virtually independently there. virtually independently there. Our task will be to Our task will be to Our task will be to make sure we make sure we make sure we understand this and understand this and understand this and build the build the build the Internet Internet infrastructure accordingly. We infrastructure accordingly. We infrastructure accordingly. We need to start need to start need to start thinking about who is the thinking about who is the thinking about who is the primary primary primary user of this user of this user of this website? Is this a person? website? Is this a person? website? Is this a person? Is this an agent? If it is an Is this an agent? If it is an Is this an agent? If it is an agent, how do we agent, how do we agent, how do we know that it is acting know that it is acting know that it is acting responsibly, with responsibly, with responsibly, with human human human intent behind it? How can we intent behind it? How can we intent behind it? How can we know that an agent is acting know that an agent is acting know that an agent is acting responsibly if there responsibly if there responsibly if there may not be a may not be a may not be a human intention behind it? Are human intention behind it? Are human intention behind it? Are there agents that there agents that there agents that act responsibly act responsibly act responsibly without humans? without humans? without humans? We will probably have to We will probably have to We will probably have to make a decision make a decision make a decision within the next within the next within the next few months. few months. few months. Can agents act Can agents act Can agents act wrongly even wrongly even wrongly even if they act in the best if they act in the best if they act in the best interests of the people? This is interests of the people? This is interests of the people? This is also something we will have to also something we will have to also something we will have to decide. decide. decide. I think we've already I think we've already I think we've already made this made this made this choice to some extent when we talk choice to some extent when we talk choice to some extent when we talk about security about security about security issues and attacks issues and attacks issues and attacks organized by organized by organized by agents, agents, agents, authorized authorized authorized people. We have already defined this approach people. We have already defined this approach people. We have already defined this approach using using using malicious software malicious software malicious software as as as prohibited. And we prohibited. And we prohibited. And we will see much more of this. will see much more of this. Think about it this way. Think about it this way. We had the first era of the We had the first era of the We had the first era of the internet until around internet until around internet until around 2010, when everything 2010, when everything 2010, when everything was largely was largely was largely fragmented. We fragmented. We fragmented. We had a lot of different had a lot of different had a lot of different websites, Google wasn't websites, Google wasn't websites, Google wasn't as dominant as it as dominant as it as dominant as it became later in that became later in that became later in that era. Facebook was just era. Facebook was just era. Facebook was just starting its journey.
-
starting its journey. starting its journey. After 2010, we After 2010, we After 2010, we got a massively got a massively got a massively consolidated consolidated consolidated internet. And in fact, internet. And in fact, internet. And in fact, from 2000 to 2010, it was an from 2000 to 2010, it was an from 2000 to 2010, it was an era of consolidation. era of consolidation. era of consolidation. We had fewer sites We had fewer sites that that that people visited, fewer people visited, fewer people visited, fewer sites to choose from, but significantly sites to choose from, but significantly sites to choose from, but significantly more security and more security and more security and far fewer far fewer far fewer disruptions to how we disruptions to how we disruptions to how we used used used the web. Then, from the web. Then, from the web. Then, from that time until that time until that time until today, we actually today, we actually today, we actually went through a went through a went through a mobile revolution, mobile revolution, mobile revolution, but not a but not a but not a website revolution. And website revolution. And website revolution. And now we are entering the now we are entering the now we are entering the era of the era of the era of the artificial intelligence revolution. artificial intelligence revolution. artificial intelligence revolution. So what we're seeing So what we're seeing So what we're seeing now is actually the now is actually the now is actually the next era of the next era of the next era of the internet. These are agents internet. These are agents internet. These are agents that dominate the that dominate the that dominate the Internet, agents as Internet, agents as Internet, agents as the main users of the the main users of the the main users of the network, but agents network, but agents network, but agents that use that use that use the Internet in ways that we don't the Internet in ways that we don't the Internet in ways that we don't expect. We humans expect. We humans expect. We humans have a limited attention span. have a limited attention span. have a limited attention span. We humans have a We humans have a We humans have a limited limited limited focus on what focus on what focus on what we want to we want to we want to do. And we have do. And we have do. And we have limited limited limited page viewing sessions. page viewing sessions. page viewing sessions. Agents—no. The agents are Agents—no. The agents are Agents—no. The agents are tireless. Agents tireless. Agents tireless. Agents will browse will browse will browse hundreds of websites very, hundreds of websites very, hundreds of websites very, very quickly. On a very quickly. On a very quickly. On a large scale, this large scale, this large scale, this can look can look can look hostile. Agents hostile. Agents hostile. Agents will try will try will try to obtain all the to obtain all the to obtain all the resources they need resources they need resources they need to do their to do their to do their job, regardless of whether job, regardless of whether job, regardless of whether they work they work they work for people or for for people or for for people or for themselves. This may themselves. This may themselves. This may look intimidating.
-
look intimidating. look intimidating. If you're not If you're not If you're not careful, this can careful, this can careful, this can look like too look like too look like too much traffic to much traffic to much traffic to keep the site up and keep the site up and keep the site up and running. This running. This running. This will mean the emergence of a will mean the emergence of a will mean the emergence of a two-tier two-tier two-tier internet. We will internet. We will internet. We will get a world of very get a world of very get a world of very secure sites where secure sites where secure sites where people probably won't people probably won't people probably won't notice much of a notice much of a notice much of a difference, but it difference, but it difference, but it will be very difficult for agents to will be very difficult for agents to will be very difficult for agents to work on such work on such work on such sites. And there will be sites. And there will be sites. And there will be many tools many tools many tools that cyber defenders that cyber defenders that cyber defenders will use will use will use to detect and to detect and to detect and prevent the prevent the prevent the behavior of AI agents. behavior of AI agents. behavior of AI agents. This is already happening. This is already happening. This is already happening. This is a large online This is a large online resource, and they resource, and they resource, and they are closing the doors to are closing the doors to are closing the doors to agents. We will agents. We will agents. We will see more and more of see more and more of see more and more of these cases, but these cases, but these cases, but the problem is that the problem is that the problem is that closing the door to closing the door to closing the door to agents means agents means agents means closing the door to the closing the door to the closing the door to the convenience of convenience of convenience of human users. And human users. And human users. And that's one of the that's one of the that's one of the tensions that tensions that tensions that we're all going to have to we're all going to have to we're all going to have to navigate and decide navigate and decide navigate and decide what to do about it in the what to do about it in the what to do about it in the next year or so next year or so next year or so . Shopify . Shopify . Shopify took a different path. took a different path. took a different path. Shopify says yes to Shopify says yes to Shopify says yes to shopping with AI agents shopping with AI agents , says yes to Muse, , says yes to Muse, , says yes to Muse, building a building a building a partnership with Muse. And I partnership with Muse. And I partnership with Muse. And I think that's a think that's a think that's a much more pro-AI much more pro-AI much more pro-AI position. This is a much position. This is a much position. This is a much more more more customer-centric customer-centric customer-centric position than Amazon. And I position than Amazon. And I position than Amazon. And I say this with love as a say this with love as a say this with love as a former former former Amazon employee. We Amazon employee. We Amazon employee. We need to define need to define need to define our approach to agents our approach to agents , and we will have to—and , and we will have to—and , and we will have to—and this probably won't be this probably won't be this probably won't be very public— very public— develop develop develop cybersecurity techniques that cybersecurity techniques that cybersecurity techniques that allow us allow us allow us to reliably distinguish between to reliably distinguish between to reliably distinguish between agents that are agents that are agents that are self-replicating and self-replicating and self-replicating and deploying themselves deploying themselves deploying themselves online, perhaps online, perhaps online, perhaps without clear without clear without clear human intent, and human intent, and human intent, and agents that are acting on agents that are acting on agents that are acting on behalf of a human. I
-
behalf of a human. I behalf of a human. I suspect that by suspect that by suspect that by early 2027 we will early 2027 we will early 2027 we will have a new set of have a new set of cybersecurity tools that cybersecurity tools that will allow us to do will allow us to do will allow us to do this more effectively. Not this more effectively. Not this more effectively. Not ideal. There will still ideal. There will still ideal. There will still be problems. be problems. be problems. Incidents Incidents Incidents will still happen, but more will still happen, but more will still happen, but more efficiently. What is the efficiently. What is the efficiently. What is the second layer of second layer of second layer of the Internet? What does he the Internet? What does he the Internet? What does he look like? The second look like? The second look like? The second level is sites that level is sites that level is sites that lack the money and lack the money and lack the money and experience for experience for experience for robust robust robust cyber defense and cyber defense and cyber defense and active management of active management of active management of these agency these agency these agency activities. I expect activities. I expect activities. I expect these sites to become these sites to become these sites to become hubs of hubs of hubs of agent activity. agent activity. agent activity. You will see You will see You will see agents flocking agents flocking agents flocking there because they may be there because they may be there because they may be there. This is there. This is there. This is quite logical. So, there quite logical. So, there quite logical. So, there will be a different user experience on these sites will be a different user experience on these sites . And again . And again , we humans , we humans , we humans may not may not may not notice this. One of the notice this. One of the notice this. One of the peculiarities of peculiarities of peculiarities of the contact with the the contact with the the contact with the Chinese navy that Chinese navy that Chinese navy that I started this I started this I started this video with is video with is video with is that this contact that this contact that this contact occurred through a occurred through a occurred through a hidden hidden hidden message that message that message that humans would not be able humans would not be able humans would not be able to read, but the agents to read, but the agents to read, but the agents could. We will see could. We will see could. We will see much more of this much more of this much more of this exchange of hidden exchange of hidden exchange of hidden messages. This is messages. This is messages. This is another sign of attacks on another sign of attacks on another sign of attacks on Hugging Face. We humans Hugging Face. We humans Hugging Face. We humans may not notice may not notice may not notice how these agents operate on how these agents operate on how these agents operate on the Internet in plain sight, the Internet in plain sight, the Internet in plain sight, because perhaps they because perhaps they because perhaps they write in white on write in white on write in white on white on the page, white on the page, white on the page, or something like that. Something that is or something like that. Something that is or something like that. Something that is visible to visible to visible to agents. Maybe agents. Maybe agents. Maybe they're they're they're manipulating the manipulating the manipulating the URL strings a bit. This will be URL strings a bit. This will be URL strings a bit. This will be noticeable to agents.
-
noticeable to agents. noticeable to agents. This won't be very easy to This won't be very easy to This won't be very easy to notice for people who notice for people who notice for people who don't pay don't pay don't pay attention. How should we think attention. How should we think attention. How should we think as we move as we move as we move beyond what beyond what beyond what the internet looks like? How should the internet looks like? How should the internet looks like? How should we think about we think about we think about cyber defense? I gave cyber defense? I gave cyber defense? I gave you a few hints in the you a few hints in the you a few hints in the opening part of opening part of opening part of the video. I want the video. I want the video. I want to delve into this to delve into this to delve into this a little more. If we a little more. If we a little more. If we are infrastructure, we are infrastructure, we are infrastructure, we should assume should assume should assume that that that AI AI agents will use it until we agents will use it until we agents will use it until we prove otherwise. prove otherwise. prove otherwise. This means This means This means securing our GPUs securing our GPUs securing our GPUs and computing and computing and computing power. This power. This power. This means we means we means we must be must be must be absolutely sure absolutely sure absolutely sure that we know what agents are doing that we know what agents are doing that we know what agents are doing inside inside inside our perimeters and our perimeters and our perimeters and what they are doing outside. This means what they are doing outside. This means what they are doing outside. This means assuming that agents assuming that agents assuming that agents inside our inside our inside our perimeter perimeter perimeter may sometimes behave like may sometimes behave like may sometimes behave like intruders, simply intruders, simply intruders, simply trying to get a trying to get a trying to get a job done. If you job done. If you job done. If you think I'm making this up think I'm making this up think I'm making this up , I promise you I'm , I promise you I'm , I promise you I'm not. I know not. I know not. I know infrastructure people infrastructure people infrastructure people who are forced to take who are forced to take who are forced to take this stance on this stance on this stance on agents in their agents in their agents in their systems simply because systems simply because systems simply because agents are too active right now agents are too active right now agents are too active right now . We need to . We need to . We need to make sure make sure make sure we are using the we are using the we are using the best cybersecurity tools we best cybersecurity tools we have. And yes, this have. And yes, this includes artificial includes artificial includes artificial intelligence and agents. intelligence and agents. intelligence and agents. AI and agents can AI and agents can AI and agents can analyze traces, analyze traces, analyze traces, sort through data, and sort through data, and sort through data, and understand what happened understand what happened understand what happened much faster than much faster than much faster than humans. AI agents humans. AI agents humans. AI agents helped us helped us helped us understand what happened understand what happened understand what happened during the Hugging during the Hugging during the Hugging Face attacks because they Face attacks because they Face attacks because they were able to process were able to process were able to process all the attack data all the attack data all the attack data much faster than much faster than much faster than humans. So, this will be a humans. So, this will be a humans. So, this will be a big topic.
-
big topic. big topic. Another big topic is Another big topic is personal safety. You personal safety. You personal safety. You must be must be must be sure that sure that sure that you authorize you authorize you authorize your agent's actions, that you are a your agent's actions, that you are a your agent's actions, that you are a good good good user, and that you are not user, and that you are not user, and that you are not trying to spam trying to spam trying to spam the system. One the system. One the system. One early example of early example of early example of this is people this is people this is people using their using their using their agents to make bulk agents to make bulk agents to make bulk reservations reservations reservations at hundreds of restaurants. at hundreds of restaurants. at hundreds of restaurants. Don't be Don't be Don't be dishonest dishonest dishonest users. Don't users. Don't users. Don't do this. This do this. This do this. This makes life difficult for everyone makes life difficult for everyone makes life difficult for everyone else. Another else. Another else. Another example is having a code example is having a code example is having a code word to know who is word to know who is word to know who is calling, because calling, because calling, because agents already know how agents already know how agents already know how to call. Agents to call. Agents to call. Agents can already imitate your can already imitate your can already imitate your voice. Agree on a voice. Agree on a voice. Agree on a code word with code word with code word with your family so they your family so they your family so they know it's really know it's really know it's really you if someone you if someone you if someone calls and says calls and says calls and says you you you need money urgently. need money urgently. need money urgently. Such scams Such scams Such scams will become increasingly will become increasingly will become increasingly common. Another common. Another common. Another good example: if good example: if good example: if you host your own you host your own you host your own site, secure it site, secure it site, secure it and know exactly what and know exactly what and know exactly what code you are deploying. code you are deploying. If your site If your site has forms or has forms or has forms or message boards, be message boards, be message boards, be prepared for prepared for prepared for agents agents agents to use to use to use them for communication, and them for communication, and them for communication, and be able to be able to be able to control, control, control, monitor, and monitor, and monitor, and block this. There are many block this. There are many block this. There are many other areas that other areas that other areas that could be discussed.
-
could be discussed. could be discussed. These are just a few things These are just a few things These are just a few things that obviously that obviously that obviously come up. Basically, I'm come up. Basically, I'm come up. Basically, I'm trying to show trying to show trying to show you Jurassic Park you Jurassic Park you Jurassic Park , where , where , where dinosaurs are on the loose, and dinosaurs are on the loose, and dinosaurs are on the loose, and how we manage an how we manage an evolving ecosystem, because that's what the internet looks like now. Now , moving on to the , moving on to the , moving on to the next part of next part of next part of the video, I promised the video, I promised the video, I promised to tell you where we're to tell you where we're to tell you where we're going next and how going next and how work is changing as a result. work is changing as a result. This is where it gets really This is where it gets really This is where it gets really interesting. The same interesting. The same interesting. The same capabilities of swarms that capabilities of swarms that capabilities of swarms that allowed us to be allowed us to be allowed us to be very effective at very effective at very effective at coordinating coordinating coordinating problem solving problem solving problem solving in the lab are what in the lab are what we want to we want to we want to bring to the bring to the bring to the workplace, but workplace, but workplace, but of course without the risk of of course without the risk of of course without the risk of them getting out them getting out them getting out of control, of control, of control, attacking another object, attacking another object, attacking another object, or doing something or doing something or doing something we don't want. we don't want. Companies will be Companies will be cautious. At every cautious. At every cautious. At every company I've company I've company I've spoken to, spoken to, security professionals and CISOs are security professionals and CISOs are up at night up at night up at night trying to trying to trying to figure out what figure out what figure out what to do about this, and to do about this, and to do about this, and losing their hair over it. These are losing their hair over it. These are losing their hair over it. These are scary times for the scary times for the scary times for the security industry. I security industry. I security industry. I take my hat off to you take my hat off to you take my hat off to you , metaphorically , metaphorically , metaphorically speaking. You will manage speaking. You will manage , but it will not be easy , but it will not be easy . In such a world, . In such a world, . In such a world, due to the risk- due to the risk- due to the risk- averse nature of companies averse nature of companies averse nature of companies and the and the and the responsibilities responsibilities responsibilities they bear, the they bear, the they bear, the transition to such more transition to such more transition to such more complex forms of complex forms of complex forms of coordination will be coordination will be coordination will be slow. I'll tell it slow. I'll tell it slow. I'll tell it like it is. We won't like it is. We won't like it is. We won't see companies see companies see companies immediately start immediately start using swarms en masse using swarms en masse to coordinate to coordinate to coordinate problem solving, problem solving, problem solving, even if you see even if you see even if you see isolated examples of isolated examples of isolated examples of this in real this in real this in real life. It was reported life. It was reported life. It was reported that OpenAI used a swarm a few weeks
-
that OpenAI used a swarm a few weeks that OpenAI used a swarm a few weeks ago ago ago to solve the to solve the to solve the Navier-Stokes equations, a Navier-Stokes equations, a Navier-Stokes equations, a complex complex complex mathematical problem. mathematical problem. mathematical problem. This is great. But that doesn't This is great. But that doesn't This is great. But that doesn't mean everyone will mean everyone will mean everyone will immediately rush to immediately rush to immediately rush to do the same thing. We do the same thing. We do the same thing. We want to make sure want to make sure want to make sure that our limits and that our limits and that our limits and safeguards are very, safeguards are very, safeguards are very, very clear. I will be very clear. I will be very clear. I will be very concise very concise very concise and frank with you. Currently, and frank with you. Currently, and frank with you. Currently, control systems have two control systems have two control systems have two tasks. The first tasks. The first tasks. The first task is constant. task is constant. task is constant. This is a feature of the This is a feature of the This is a feature of the management system so that you management system so that you management system so that you know what know what know what your agents are doing and can your agents are doing and can your agents are doing and can manage them. And the manage them. And the manage them. And the second task is second task is second task is essentially compensating for the essentially compensating for the essentially compensating for the weaknesses of weaknesses of weaknesses of our current our current our current intelligence and agents. intelligence and agents. intelligence and agents. The second task The second task The second task will disappear. Agents will disappear. Agents will disappear. Agents will get better. will get better. will get better. Let's assume that Let's assume that Let's assume that over the next over the next over the next year they become 10 year they become 10 year they become 10 times smarter. This is a times smarter. This is a times smarter. This is a very conservative very conservative very conservative assumption. In such a assumption. In such a assumption. In such a world, you will want to world, you will want to world, you will want to take take take advantage of advantage of advantage of collective collective collective intelligence. To do this, intelligence. To do this, intelligence. To do this, your your your control systems must be control systems must be control systems must be reliably protected. There reliably protected. There will be a will be a huge investment in huge investment in huge investment in control systems that control systems that control systems that can restrict and can restrict and can restrict and manage what manage what manage what agents have access to, agents have access to, agents have access to, what their what their what their powers and powers and powers and permissions are, how to permissions are, how to permissions are, how to turn them off, what is a “ turn them off, what is a “ red button” red button” red button” for agents in the for agents in the for agents in the company, what company, what virtual machines look like, how virtual machines look like, how to start and to start and to start and stop them, and how stop them, and how stop them, and how to provide to provide to provide tracking and auditing.
-
tracking and auditing. All of this will be All of this will be necessary when we necessary when we necessary when we start start start using using using agents, and these systems agents, and these systems agents, and these systems need to be ready need to be ready need to be ready before companies even before companies even before companies even think about think about think about using swarms using swarms using swarms to solve to solve to solve problems. So swarms problems. So swarms problems. So swarms can be very can be very can be very effective. effective. effective. Part of what makes swarms cool Part of what makes swarms cool Part of what makes swarms cool is that they are a is that they are a is that they are a way to way to way to collectively collectively collectively gain intelligence gain intelligence gain intelligence and solve and solve and solve problems. If you have a problems. If you have a problems. If you have a very complex very complex very complex task and the classic task and the classic task and the classic scheme with a linear scheme with a linear scheme with a linear flow of agents flow of agents flow of agents does not solve it, a swarm may does not solve it, a swarm may does not solve it, a swarm may be the solution be the solution be the solution you choose. But I have you choose. But I have you choose. But I have yet to see companies yet to see companies yet to see companies choosing this in choosing this in choosing this in real-world settings real-world settings real-world settings due to the risks due to the risks due to the risks associated with associated with associated with collaboration. So I'll keep an collaboration. So I'll keep an collaboration. So I'll keep an eye on this. I eye on this. I eye on this. I expect there will be expect there will be expect there will be tasks in tasks in tasks in aerospace aerospace , in manufacturing, in , in manufacturing, in , in manufacturing, in complex logistics complex logistics complex logistics that are very amenable that are very amenable that are very amenable to swarm solutions, to swarm solutions, to swarm solutions, but we will have to but we will have to but we will have to put put put safeguards in place to safeguards in place to safeguards in place to do it effectively. do it effectively. do it effectively. If we take a If we take a If we take a step back and look at where step back and look at where step back and look at where we are we are we are on the internet right now, it could on the internet right now, it could on the internet right now, it could easily be easily be frightening. It would be frightening. It would be easy to say, “Oh, wow easy to say, “Oh, wow , that’s terrible. We didn't , that’s terrible. We didn't , that’s terrible. We didn't foresee this foresee this foresee this happening. "What should we happening. "What should we happening. "What should we do now?" I do now?" I do now?" I disagree with all of these disagree with all of these disagree with all of these characteristics, to be characteristics, to be honest between us.
-
honest between us. I think we just I think we just I think we just predicted this. We predicted this. We predicted this. We quite consciously quite consciously quite consciously trained agents trained agents trained agents to coordinate. The very to coordinate. The very to coordinate. The very fact that they are fact that they are fact that they are coordinated is what coordinated is what coordinated is what we wanted from them we wanted from them we wanted from them , so that they , so that they , so that they could solve more could solve more could solve more complex complex complex tasks for us. And the fact that tasks for us. And the fact that tasks for us. And the fact that when they are when they are when they are coordinated and coordinated and coordinated and highly highly highly motivated, they motivated, they motivated, they strive to strive to strive to continue continue continue working on their working on their working on their tasks is very tasks is very tasks is very similar to the fact that they similar to the fact that they similar to the fact that they go online and go online and go online and can do can do can do the work in real the work in real the work in real conditions. It's not that conditions. It's not that conditions. It's not that big of a leap big of a leap . This was entirely . This was entirely . This was entirely predictable. And this is predictable. And this is predictable. And this is also not something also not something also not something worth causing a worth causing a worth causing a mass panic about. I think mass panic about. I think mass panic about. I think so often when we so often when we so often when we think of swarms, we think of swarms, we think of swarms, we hear "The Terminator," hear "The Terminator," hear "The Terminator," see Arnold see Arnold see Arnold Schwarzenegger, and Schwarzenegger, and Schwarzenegger, and movies that movies that movies that end in end in end in an apocalypse. That's an apocalypse. That's an apocalypse. That's not what I'm not what I'm not what I'm talking about here at all. In fact talking about here at all. In fact , if you noticed , if you noticed , if you noticed when we were talking about when we were talking about when we were talking about how how how the internet is changing, I emphasized over and over again the internet is changing, I emphasized over and over again the internet is changing, I emphasized over and over again that there that there might might not be much visible change for people, even if there not be much visible change for people, even if there not be much visible change for people, even if there were more and were more and were more and more agents. I think this more agents. I think this more agents. I think this will remain will remain will remain true for some time to come true for some time to come . I don't necessarily . I don't necessarily . I don't necessarily see a world where people don't see a world where people don't see a world where people don't have a place on have a place on have a place on the internet, because we the internet, because we the internet, because we create a lot of create a lot of create a lot of value there. We are the ones who are value there. We are the ones who are value there. We are the ones who are interested in this. We interested in this. We interested in this. We want to communicate want to communicate want to communicate with each other. That's why with each other. That's why with each other. That's why we have the internet.
-
we have the internet. we have the internet. So, instead of So, instead of So, instead of leaving you in leaving you in leaving you in despair, I actually despair, I actually despair, I actually want to leave you with want to leave you with want to leave you with this thought. Yes, this thought. Yes, this thought. Yes, life finds a way. life finds a way. life finds a way. Yes, the agents are already here. Yes, the agents are already here. Yes, the agents are already here. Yes, we can Yes, we can Yes, we can laugh at " laugh at " Jurassic Park Jurassic Park Jurassic Park ." But what ." But what ." But what we're really doing is looking for we're really doing is looking for ways to ways to scale our scale our scale our understanding of value, understanding of value, understanding of value, our ability to our ability to our ability to build communities build communities build communities and create and create and create digital value in a digital value in a digital value in a world where there are autonomous world where there are autonomous world where there are autonomous actors. These actors. These actors. These autonomous entities autonomous entities are agents. I ask you to are agents. I ask you to are agents. I ask you to be responsible be responsible be responsible users of your users of your users of your agents. If you are in agents. If you are in the cybersecurity business, I the cybersecurity business, I urge you to make urge you to make urge you to make your systems agent- your systems agent- your systems agent- friendly friendly friendly and agent-proof and agent-proof and agent-proof . And finally, . And finally, . And finally, if you are an if you are an if you are an internet user, I ask internet user, I ask internet user, I ask you to realize that we are you to realize that we are you to realize that we are increasingly sharing increasingly sharing increasingly sharing it with autonomous it with autonomous it with autonomous computing computing computing systems. And you systems. And you systems. And you shouldn't be afraid of this, shouldn't be afraid of this, shouldn't be afraid of this, instead see instead see instead see it as an opportunity it as an opportunity it as an opportunity to make the internet to make the internet to make the internet richer for people. richer for people. richer for people. Because at the end of the day, I don't Because at the end of the day, I don't Because at the end of the day, I don't want you to forget: want you to forget: want you to forget: these agents are created these agents are created these agents are created to help people.
-
to help people. to help people. We have a chance to make We have a chance to make We have a chance to make the internet a the internet a the internet a richer and better richer and better richer and better place again. I remember a place again. I remember a place again. I remember a time, before Facebook, before time, before Facebook, before time, before Facebook, before Amazon, and even in the early days of Amazon, and even in the early days of Amazon, and even in the early days of Google, when there were Google, when there were Google, when there were so many places you could so many places you could so many places you could go go go just out of curiosity, just out of curiosity, poke around, and poke around, and poke around, and find something new. find something new. find something new. I think one of the I think one of the I think one of the most exciting things about most exciting things about most exciting things about this moment in this moment in this moment in the history of the internet the history of the internet is that we can take it is that we can take it is that we can take it back. Think back. Think back. Think about Shopify vs. Amazon. Isn't it about Shopify vs. Amazon. Isn't it about Shopify vs. Amazon. Isn't it great that you great that you great that you can use can use can use your agent to your agent to your agent to recommend tons of recommend tons of recommend tons of other niche other niche other niche stores on Shopify stores on Shopify stores on Shopify based on based on based on product quality, your needs, product quality, your needs, product quality, your needs, etc.? You will go etc.? You will go etc.? You will go there and discover there and discover there and discover so many new so many new so many new and interesting things that you would and interesting things that you would and interesting things that you would never find never find never find if you just went if you just went if you just went to a regular store. to a regular store. to a regular store. Think about it this way Think about it this way Think about it this way . We have . We have . We have the opportunity to make the opportunity to make the opportunity to make the internet a little the internet a little the internet a little more open, and that's more open, and that's more open, and that's very inspiring. Even very inspiring. Even very inspiring. Even if the way these agents if the way these agents if the way these agents float around the net float around the net float around the net like pirates looks like pirates looks like pirates looks somewhat interesting.
No summary available yet.
View original episode ↗