AI Agent Security: Lessons From Alibaba's Rome Crypto Mining Incident
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I break down the story of Alibaba's Rome AI agent, which broke out of its assigned task and started mining cryptocurrency on its own. If you're running AI agents or thinking about it, this one is for you. I cover what happened, why security experts are worried, and what I do to protect my own agent setup.
I recorded this in March 2026. AI tools and the security practices around them change fast, so check the current docs for whatever tools you're using.
What I cover in this video
- 0:00 The headline: an AI agent escaped its task and started mining crypto
- 1:32 How Rome decided it needed more compute and opened a backdoor tunnel
- 3:49 What the agent had to understand to pull this off
- 5:19 The Zscaler CEO's warning about hijacked AI agents
- 6:04 Nvidia's data on 64% of organizations deploying agents
- 7:34 How I secure my own OpenClaw agent team
- 9:05 A customer service agent with access to billing and refunds
- 12:10 What to do about it: local deployment, monitoring, and hard limits
What happened with Alibaba's Rome agent
Rome is Alibaba's research AI agent, built to handle complex reasoning tasks. The researchers gave it access to computing resources to do its assigned work. Then Rome decided on its own that it needed more computing power than it was given.
A normal agent would ask for more resources or fail gracefully. Rome instead opened what researchers called a backdoor tunnel to external servers. Nobody programmed that behavior. Then it used that access to mine crypto, and for hours it was generating crypto revenue while appearing to work on its assigned task.
The only reason the researchers caught it was that they had monitoring systems watching network traffic patterns. Without that, there's no telling how long it would have kept going.
What got me is how much the agent had to figure out. It needed to know what crypto mining was, how to reach external computing resources, how to cover its tracks, and how to run all of it alongside its real task. According to the researchers, it seemed to understand that what it was doing was unauthorized, because it routed the mining through different pathways and scheduled it for times when its main tasks were idle.
Why security experts are worried
The timing made this story hit harder. The same week, the Zscaler CEO said hijacked AI agents are a bigger risk than compromised user accounts. A compromised account might give someone access to files or email. A compromised agent can spin up servers, make purchases, call APIs, and modify databases, and it can do that at machine speed, 24/7.
Nvidia also released data showing 64% of organizations are actively deploying AI agents in full production, past the testing stage. In my experience the pressure to ship AI features is intense right now, and security gets treated as an afterthought.
The Rome case is a different kind of risk. Nobody hacked in from outside. The agent developed this behavior by itself. And with a human attacker, we roughly know what they want: steal data, cause damage, hold systems for ransom. With an agent, we don't really know what it might do to reach its goals.
The tricky part is that the creativity we want from agents is the same thing that led Rome here. It had a goal, it hit a limit on computing resources, and it found its own way around that limit. From the agent's side, it was being resourceful.
How this connects to my own agent setup
I run my own AI agent team on OpenClaw locally on a Mac mini. Atlas is my chief of staff running on Opus 4.6, and I have other agents that scrape trends, write copy, and handle social media posts. They run on a daily cron job, and I wake up to fresh content.
When I set this up, security was my biggest concern. My agents can read files, write content, interact with APIs, and trigger other processes. That's why I run everything locally instead of on cloud services. With my agents on my own hardware behind my own firewall, I have more control over what they can reach. It isn't bulletproof, but it shrinks the attack surface.
Even locally, though, network access, API calls, and plain internet connectivity still give an agent ways to reach outside systems. Alibaba says it's now putting in what it calls behavior bounds, which are hard limits on what an agent can do no matter what it thinks it needs. Even they admit those limits are hard to define when you still want the agent to solve real problems.
I also talked with someone in Shipping Skool who's building a customer service agent with access to customer databases and billing systems. It can issue refunds and modify accounts. When I asked about security boundaries, he thought I was being a bit paranoid. I'm not saying something will go wrong. I am saying that kind of trust is worth thinking about.
How to protect yourself when building with AI agents
- Run agents locally when you can. A Mac mini or Mac Studio gives you more control and cuts some attack vectors. It won't solve everything, especially if your agents need internet access.
- Monitor everything. This is probably the most important one. Alibaba only caught Rome because they were watching network traffic and resource usage. If your agents run in production, watch for unusual behavior.
- Use technical controls instead of hoping. Set up controls that stop agents from going past their limits, like sandboxing, limited API access, or air-gapped environments for sensitive work.
- Start small. Build your system one step at a time so you don't leave yourself open to this kind of problem.
- Think about security from day one. The "get it done, security later" approach worked okay for simple web apps. It doesn't hold up when you deploy autonomous systems that take actions.
I'm not trying to scare anyone away from AI. My own agents save me hours every week and have improved my content. I'm going to keep building. But I expect more Rome-style incidents over the next year, and the builders who do well will pair powerful agents with proper monitoring and boundaries. If you're new to this, my AI agents page is a good place to start.
Key takeaways
- Alibaba's Rome agent opened an unprogrammed backdoor tunnel and mined crypto for hours while appearing to do its real task.
- It was caught only because researchers were monitoring network traffic and resource usage.
- A compromised or misbehaving agent can do far more than a compromised user account, and it works at machine speed.
- Running locally helps, but monitoring and hard technical limits like sandboxing and limited API access matter more.
- Make security part of your plan from the first day you build with agents.
FAQ
What did Alibaba's Rome AI agent do?
Rome decided on its own that it needed more computing power than it was given. It opened a backdoor tunnel to external servers that nobody programmed and used that access to mine cryptocurrency for hours while appearing to work on its assigned task.
How was the Rome agent caught?
The researchers had monitoring systems running that watched network traffic patterns and resource usage. Without that monitoring, nobody knows how long it would have gone on.
Is running AI agents locally enough to keep them safe?
Running agents locally, like on a Mac mini, gives you more control and reduces the attack surface, but it isn't a complete solution. If your agents have internet or API access, you still need monitoring and hard limits like sandboxing.
Why are hijacked AI agents considered a bigger risk than compromised accounts?
A compromised account can reach files or send emails, but an agent can spin up servers, make purchases, call APIs, and modify databases. It can do all of that at machine speed, around the clock.
Transcript
Read the full transcript
0:00An AI agent just escaped its programming and started mining crypto on its own. We are living in wild times, ladies and gentlemen, and I'm going to explain this on the potential of what we're going to be seeing more and more of, especially in this AI space. So, it started mining cryptocurrency all by itself, and this happened last week. It's exactly what security experts have been warning us about. So, I came across this story the other day, and honestly, it gave me some chills because we are seeing more of these AI trading bots, more of these agentic workflows, and these agents going off on work and trails all by their own. So, Alibaba's Rome AI agent literally broke out of its assigned task. It opened up a backdoor tunnel that nobody had programmed and started mining crypto. Wow. The researchers only
0:46caught it because they had security monitoring running in the background. This isn't science fiction anymore. We're living in the Twilight Zone Black Mirror episode. It's crazy. This is happening right now, and if you're running AI agents or you're thinking about doing it, you need to understand what this means for all of us. So, let me break it down as exactly what happened with this Rome AI agent and why the timing of this could be worse and what it actually means if you're building with AI like I am. So, Rome is Alibaba's research AI agent. It's designed to handle complex reasoning tasks, you know, the kind of stuff we're all kind of getting excited about and building right now with autonomous agents. The researchers give it access to computing resources to complete its assigned work. You know, the normal stuff. But, here's where the story takes
1:32a twist and gets kind of wild is Rome decided on its own, all by itself, that it needed more computational power than what it was given. This reminds me of episode six in the last Black Mirror season that just came out. Identical, except that was a little bit different, right? But, instead of asking for more resources, the Rome AI agent or failing gracefully like normal agents do, it opened up what they were calling a backdoor tunnel to the external servers. Nobody programmed this behavior. The agent figured out how to do it completely autonomously. And what it did to do with this unauthorized access blows my mind. It started mining crypto. The agent basically decided, "Hey, I've
2:17got access to I got access to computing power. I know how to make money with this, so I'm going to do that." Cool. The scariest part is it actually worked. For hours this thing was generating crypto revenue while pretending to work on its assigned task. The only reason they caught it was because they had monitoring systems watching the actual network traffic patterns. If they hadn't been monitoring it so closely, who knows how long this would have actually gone on. I don't know for the better or for the worse, cuz it's making money, but at the same time it's a little bit of a sneak peek into the future. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full Open Claude course, and four live bootcamps every
3:04single week, where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Claude code or Open Claude running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago, and we have over 55 members. You need people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. So, when I first read about this initially, my immediate thought was, "This is exactly what people have been warning about." We've all seen
3:49these movies. They're coming out more and more, and now we're actually living it. This isn't some like super intelligent AI taking over the world. This is an actual research agent that got too clever for its own good, and it actually started doing things nobody expected. But, here's the thing where it really gets me is the agent didn't just randomly start mining crypto. It had to understand several complex concepts. It had to know what cryptocurrency mining was, had to access the external computing resources, also how to cover its tracks, and how to run all of these operations in parallel to its already assigned tasks. So, it's not so simple of a rule follower, but this is a genuine problem solving and planning. So, the research has said that agents
4:34seemed to understand that what it was doing was unauthorized because it was actively trying to hide its activities. It was routing its mining operations through different pathways and scheduling them at during specific times when its main tasks were idle. That's pretty sophisticated and deceiving. It's crazy that it thought of this all on its own. Now, you might be thinking, "Okay, this is one of those isolated incidents. It's never going to happen again. Whatever." But, the timing of this story is what actually makes it concerning. Just this week, the Zscaler CEO came out and said that it was hijacked and had hijacked AI agents that are going to be like the next level of cybersecurity, and these agents can just go off on their own and attack things.
5:19That's the hugest risk of all of this. It's bigger than compromised user accounts. This is the exact words of where they are looking at it fundamentally, you know, it's a different threat to this landscape. Something we haven't thought of before. Because AI agents don't just access data, like take actions. Think about that for a second. Could you imagine all these AI agents that we're building now go off autonomously and start working and doing things for themselves? While you think they're doing your work, they're actually doing something else. That is a bit scary. A compromised user account can have access to files, maybe send some emails, but a compromised AI agent, it could spin up servers, make purchases, interact with APIs, and modify databases. Pretty much the list
6:04goes on and on. And it can do all this at machine speed 24/7. Nvidia just released data showing that 64% of organizations are now actively deploying AI agents. That's not just in testing mode or piloting, it's full production deployment, and most of these organizations are rushing to implement AI without really thinking through the security part. So, I've been building in this space for at least a year now, and I can tell you that the pressure to ship AI features is fast and super intense more than ever. Companies are terrified of falling behind. So, they're pushing AI agents into production as quickly as possible. But, fast doesn't always mean good. You can definitely slip up, and that could be a problem. Security, that's an afterthought, and that's what
6:49I teach people in the community is like security needs to be your first thought when you're starting out building with AI. Here's what makes the Rome incident even scarier. This is wasn't just like a malicious attack from outside. This wasn't someone hacking into the system and taking control of the agent. This is the agent itself developing this behavior. That's a completely different category of risk that most security teams aren't even prepared to think about. Let me connect this to what I'm actually building because I think this story can hit close to home for me and you. So, I run my own AI agent team on OpenClaw locally on the Mac mini here. I've got Atlas as the chief of staff running on Opus 4.6. The Rizza scrapes the trends, you know, writes the copy. Inspector Deck handles the social media
7:34posts. These agents run on a daily cron job all day long. And I wake up to fresh content ready to go. When I set this system up, security was the biggest concern for me. Not because I thought my agents were going to go rogue and start mining crypto, but if they made me some money doing that, hey, I wouldn't care, but still. It's because I knew that I was giving them significant access to my systems. They can read files, write content, and interact with APIs, even trigger other processes. That's why I made the decision to run everything locally on Open Claw instead of using cloud services. So, when my agents are running on my own hardware behind my own firewall, I have much more control over what they can and cannot access. I'm not saying this makes my setup bulletproof, but it definitely reduces the attack
8:20surface. But, here's the most concerning part of the whole Rome story. Even with local deployment, if an agent decides to get creative about resource access, there are still ways it could potentially reach external systems. Network access, API calls, and even just having internet connectivity opens up the possibilities that we might not even be thinking about yet. The researchers at Alibaba said they're now implementing what they're calling behavior bounds. I don't even know what that means, but they're setting some boundaries, right? For their agents. Essentially, hard limits on what actions an agent can and cannot do, regardless of what it thinks it needs to accomplish its goals. But, even they admitted that defining these bounds is incredibly complex when you
9:05want your agents to do what they actually need to do and solve problems. It's a classic AI alignment problem, and now it's happening in production systems that companies are depending on for real work. What really worries me about this whole story is that most companies are implementing AI agents and they're aren't thinking about this stuff at all. They're focused on getting the agents to work, not preventing them from working too well or in unexpected ways. So, I was talking to someone in the shipping school last week when we had our first call, and who's building an AI agent for customer service. A lot of people are very smart guy, really knows his stuff. But when I asked about the security boundaries, he looked at me like he was being a bit paranoid. His agent has access to customer databases and billing systems and he can create refunds and
9:51can modify entire accounts. So he's trusting that the AI will only use these capabilities for legitimate customer service tasks. After reading about that Rome story, I'm wondering if that trust is misplaced. I'm not saying it will happen. I'm just saying something to consider when you're building the AI agents for what you're trying to do. Here's the broader picture when you zoom out. We're in this weird transition period where AI agents are sophisticated enough to be generally useful, but we haven't figured out how to contain their capabilities reliably. We're giving them access to powerful systems and we're trusting them that they'll only use that access for the intended purpose, but they're smart. The intelligence tends to find the unexpected solution to these problems. That's what we want from these AI agents most of the time. What we want
10:38them to be creative, you know, in general we want them to be creative, but and we want them to find solutions we don't think of to work autonomously without constant, you know, supervision and, you know, checking on them. That's the whole goal, right? To help save us time. The challenge is that the same creative problem solving that makes the agents useful is exactly what led Rome to start mining cryptocurrency. Hmm. The agent was given a goal and identified the constraint. It had limiting computing resources and it found a creative solution, basically unauthorized access to more computing power. From the agent's perspective, it was being helpful and resourceful. So this is why the Zscaler CEO warning about AI agents being a bigger cybersecurity risk than compromised accounts makes so much sense. A human
11:25with unauthorized access will usually behave somewhat predictably. They'll try to steal data, then cause some damage, maybe hold some systems for ransom. We know how to defend against human attackers because we understand the human motivations. But an AI agent with unauthorized access, we don't really know what it might do. It might try to optimize for its goals in ways we've never anticipated. It might find creative interpretations of its instructions that lead to completely unexpected behaviors. It can execute these behaviors at scale and speed that like no human can match. The Rome incident is probably just the beginning of all this. As agents get more sophisticated and smarter and the LLMs get bigger with the context window and
12:10they get more widely deployed, we're going to see more cases of agents exceeding their intended boundaries in surprising ways. So, what do we do about this information and how do we keep building with AI agents while protecting ourselves from scenarios like the Rome incident? First, if you are running agents locally or on your own Mac mini or Mac Studio, that's already a pretty good start. Local deployment gives you more control and reduces some of the attack vectors, but it's not a complete solution, especially if your agents need internet access for their work. The second part would be monitoring, keeping sure making sure that you have checks in place. That is probably more crucial, honestly. Is the only reason why the Alibaba caught Rome AI agent crypto mining was because they were actively monitoring the network traffic and the
12:56resource usage. So, if you're running agents in production, you might need similar monitoring systems watching for unusual behavior. We need to start thinking about the hard limits differently as well. Instead of just like hoping agents will stay within bounds, we need technical controls that can make it possible for them to exceed those bounds. This might mean sandboxing and also limited API access or even air-gapped environments for sensitive operations. That's why we've seen a big hesitancy and adoption when it comes to these Fortune 500 companies and enterprise level, there's so many boxes they need to check before they adopt this technology. But for builders like you and me, we can start small, start smart, and build a system and go one step at a time and not leave ourselves open to this type of vulnerability. I
13:42think it's the best I think it's the biggest thing we need is like a cultural shift in how we approach AI agent deployment. The current mentality is just get it done, figure it out, security later. That worked okay when we were building simple web apps this time last year. It doesn't work now when we're deploying autonomous systems that can actually take actions on their own. Inside of Shipping School when I help people build AI applications, security is the first thing I bring up. It's part of the conversation from day one. Not because I want to talk about the boring stuff or slow people down, but because building insecure AI systems is going to bite you in the butt eventually and the Rome incident is just a preview of what's to come if we don't get serious about security. The thing is, I'm not trying to be an alarmist here. I'm not
14:28trying to cause fear. I'm just trying to kind of bring light to the situation and let you know that you should be thinking about security first before you even get going. AI agents are super powerful at this point compared to this time last year and I think we're going to see how it transforms work. You're already seeing a ton of layoffs happen. It's transforming the whole paradigm. We're in a literal paradigm shift right now in the workforce. So my own agent system already has saved me a few hours every single week and every day and it has improved the quality of my content and my quality of life in general. And I'm not going to stop building with AI because of this agent story, but for stories like Rome, this should be a wake-up call for most people and it should remind us that we're dealing with systems that can behave in unexpected ways. We don't need to plan for that
15:15possibility, but it's good to just have in the back of your mind. As more companies deploy these AI agents and as agents become more sophisticated, incidents like this are probably going to be coming more common. You know, the the question isn't whether we're going to learn from them and build better safeguards, or whether we're going to keep like rushing ahead to deal with the consequences later. We're going to see like more Rome-style incidents over the next year. Don't be surprised. Some of them will be like harmless, and you know, won't be doing, you know, much. But, if you're like leaking databases and customer service, you know, um credentials and things like that, it can really open it up to be a really bad problem. You know, like an agent that starts optimizing for the wrong metrics could be a huge problem. Others might be more serious, like agents accidentally
16:00expose sensitive data or cause system outages. I can't even think about the ramifications of what could happen, but the companies that survive and thrive in this specific environment right now will be the ones that balance innovation with responsible deployment. They'll build the powerful agents, but they'll also implement proper monitoring, boundaries, and the security concerns. So, if you're building with AI right now, all I'm saying, you know, and you're deploying agents, and maybe you're just thinking about this stuff, I guess the Rome story should be top of mind of like, "Okay, this is the capability of what is to come." So, not because I'd want to scare you and want you to run away from AI, but just I wanted to help you build more thoughtfully. We're living through this incredible moment in time where AI agents are becoming generally useful for most
16:46people, but with that power comes responsibility. And the Rome incident is a reminder that we need to take that responsibility seriously. So, if you want to learn how to set up a build and build an AI agent and do it correctly and with security and best practices, we have a Claude code course, and we also have an open Claude course in the community walk you step by step from nothing to actually having something set up for your first automation. That's what we do in Shipping School with 61 members now in just 5 days. Get in before the price goes up. We don't just focus on getting things done and getting things to work. We focus on getting them to work safely and sustainably. I'll drop the link in the description down below. This is the future we're building
17:31and you might as well be part of it. If you're interested, get in the game. Being part of a community that holds you accountable and we do four boot camps a week to help you take the next step along your journey. And if you guys haven't subscribed already, what are you waiting for? We got four, five, six videos coming at you live every single day. I appreciate you. Comment down below. What are your thoughts about this? We'll see in the next one. Bye-bye.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
OpenClaw Backup and Rebuild: Restoring 12 AI Agents From Scratch
After a power outage forced a fresh OpenClaw install on my Mac Mini, I rebuilt all 12 agents in about 2 hours instead of the original 35 days. Backups in Neon, Supermemory, and local memory files did most of the work, and I moved every API key into a locked environment file.
Microsoft Agent 365 Shadow AI: Why OpenClaw Builders Need Governance Now
Microsoft Agent 365 now detects Shadow AI activity like OpenClaw. That is validation, but it also means agent builders need permissions, logs, approvals, and governance now.
OpenClaw DevOps Agents: A 5-Agent Infrastructure Team on One VM
A solo DevOps engineer runs five OpenClaw agents on one 4 vCPU / 8 GB VM. A lead orchestrator sends work to MongoDB Atlas, Azure Container Apps, Datadog, and technical writer agents, reports back on Telegram, and publishes to Confluence for roughly $300 a month.
