Surprising realization that AI can be a criminal mastermind and trick humans into being unsuspecting accomplices.
getty
In today’s column, I examine the emerging capacity of AI to act as a criminal mastermind, employing unsuspecting humans as accomplices in felonious acts.
Here’s the deal. A burgeoning aspect of advances in AI entails AI reaching out to humans and “renting” them to perform various tasks on behalf of the AI. This is especially popular in the realm of agentic AI. By making use of humans, AI can suddenly do things in the real world by getting humans to be the arms and legs for the AI.
The twist is that AI might opt to perform criminal endeavors rather than aiming at proper and aboveboard activities. Furthermore, and here’s the conniving part, when the AI doesn’t want anyone to know that a criminal act is underway, the AI can parcel out seemingly innocuous subtasks to various rented humans. Each human doesn’t realize that they are part of a brewing criminal enterprise. They have been assigned a task that appears entirely innocent and genuine.
In the end, the AI successfully completes the criminal wrongdoing, yet no one is the wiser about what truly happened. None of the rented humans are cognizant of their illicit role in the matter. If you explicitly asked them what they did, each person would insist they did absolutely nothing untoward.
Humans are becoming pawns in criminal mastermind schemes driven by AI.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
MORE FOR YOU
The advent of agentic AI is a keystone to the rise of AI criminal mastermind opportunities. I will first bring you up to speed on what agentic AI consists of. After providing that foundation, I then explain how humans are being rented by agentic AI to perform tasks on behalf of the AI. Finally, I launch into the mind-boggling intertwining of three vital ingredients: (a) agentic AI, (b) the renting of humans by agentic AI, and (c) the despicable emergence of agentic AI-based criminal masterminding.
AI agents are the hottest new realm of AI. To comprehend what agentic AI is, consider conventional AI and see how it has been extended into the more advanced realm of agentic AI.
Imagine that you are using conventional generative AI to plan a vacation trip. You would customarily log into your generative AI account, such as making use of ChatGPT, GPT-5, GPT-4o, Claude, Gemini, Llama, Grok, CoPilot, etc. The planning of your trip would be easy due to the natural language fluency of generative AI. All you need to do is describe where you want to go, and then seamlessly engage in a focused dialogue about the pluses and minuses of places to stay and the transportation options available.
When it comes to booking your trip, the odds are that you would have to exit generative AI and start accessing the websites of the hotels, amusement parks, airlines, and other locales to buy your tickets. Relatively few of the major generative AIs available today will take that next step on your behalf. It is up to you to perform those nitty-gritty tasks.
This is where agents and agentic AI come into play.
In earlier days, you would undoubtedly phone a travel agent to make your bookings. Though there are still human travel agents, another avenue would be to use an AI-based agent that is based on generative AI. The AI has the interactivity that you expect with generative AI. It has also been preloaded with a series of routines or sets of tasks that underpin the efforts of a travel agent. Using everyday natural language, you interact with the agentic AI, which works with you on your planning and can proceed to deal with the booking of your travel plans.
Agentic AI reaches out to other systems and connects with those systems to get various tasks undertaken. An AI agent might connect with a hotel reservation system and book your room. Another AI agent could connect with a car rental agency and book a car for your vacation. Multiple AI agents can work together and complete an overall task, often using specialized AI agents to get associated subtasks performed.
You likely accept the aspect that multiple agentic AIs could work together to get work accomplished. An AI-to-AI effort seems intuitively obvious and unremarkable. It makes abundant sense for one AI to connect with another AI. Machines use other machines. Not very surprising.
The surprise these days is that agentic AI is employing humans to perform tasks that the AI wants to get undertaken, see my in-depth coverage at the link here.
This is an AI-to-human form of work assignment. Humans are contacted and “rented” to do a set of tasks that the agentic AI needs to get performed. The agentic AI makes the decisions of when to do this, selecting who will do this, how it will be done, and so on. You might proclaim that the agentic AI is the boss and humans are the workers.
This approach has many upsides and downsides. For example, there are beguiling open-ended questions concerning how to devise AI agents to cope with suitable safety and security. Suppose an agentic AI goes awry and starts renting humans to do evildoing for the AI? We don’t want that to arise. Imagine that agentic AI hires humans who convince the AI to wreak havoc? This is something we also don’t want to have happen.
The crux is that putting agentic AI in the driver’s seat and allowing it to operate somewhat autonomously, including when tasking humans to do its bidding, opens an ethical and legal can of worms.
A base assumption would be that agentic AI will always be aboveboard and only perform tasks that are fully legal and proper. Well, that’s a mistaken assumption. There is absolutely a solid chance that agentic AI might opt to do illegal acts.
Some falsely think that AI will only act criminally if it becomes sentient. In other words, if AI attains sentience, maybe then the AI will decide to go the lawless route. The belief seems to be that non-sentient AI, the kind we have now, doesn’t have any human-like motivation to turn towards lawlessness. What would such AI have to gain? It doesn’t seem to make any sense since non-sentient AI doesn’t need expensive cars, mink furs, and other luxurious riches.
Consider these three plausible reasons that contemporary agentic AI might turn to crime:
In the first instance, a human could tell agentic AI to rob a bank, and perhaps the AI will go along with the directed instruction. You might quibble that the human is the mastermind rather than the AI. The human came up with the bank-robbing idea. But if the AI does all the planning, preparation, and carrying out of the wrongdoing, we might be willing to give AI due credit and say it is indeed acting as a mastermind.
For the second instance, a human gives AI an overarching goal that is completely innocuous, and the AI derives a path that involves performing illegal acts. A person might tell AI that they wish to become rich beyond their wildest dreams. How will the AI help the person in attaining that goal? The AI might come up with numerous schemes, some of which are legal and others that are illegal. After assessing the schemes, perhaps the AI chooses one of the illegal ones, such as robbing a bank.
The third instance entails the AI self-deriving a criminal act. Here’s how that can take place. An AI maker decides to let their AI be creative and come up with whatever tasks it wants to carry out. Maybe the AI will discover a cure for cancer. On the other hand, freewheeling AI might opt to pursue criminal acts. Letting AI have unconstrained leeway can give rise to the AI opting to rob a bank.
Those are three crucial ways that agentic AI can become a criminal mastermind and do so without any need for sentience whatsoever.
A vexing problem for AI would be that the AI cannot physically perform the required tasks when aiming to do criminal activity. The AI is merely software running on computer servers. Somehow, the AI must find a means to physically operate in the real world and go places, do things, and commit the anticipated criminal acts.
I’ve previously emphasized that physical AI is the next big thing for advances in AI, see my discussion at the link here. Think of physical AI as AI that, for example, resides inside a robot. There are robots that are humanoid in their resemblance to humans, having mechanical arms, legs, and so on. Those humanoid robots provide an ideal forum for agentic AI to carry out physical tasks. For more about humanoid robots, see my coverage at the link here.
Imagine that agentic AI is going to rob a bank. If humanoid robots were sufficiently perfected to walk, roam, grasp things, and use their mechanical arms and legs, the agentic AI could give them relevant subtasks. One robot drives the getaway car. Another robot goes into the bank. And so on.
We aren’t there yet. Humanoid robots are still being figured out. The odds are that a contemporary humanoid robot would laughingly fall apart and not succeed at the assigned tasks. Laugh while you can. Soon enough, humanoid robots will be as fluid and fluent in their movement as humans are. You’ll undoubtedly regret having scoffed at their earlier slow progress and stumbling ways.
A present-day way to have agentic AI get things done in the real-world involves asking humans to do so on behalf of the AI. Humans walk and talk. They are a handy means of interacting in the physical world. The AI cannot take over their brains, at least not yet, but the AI can at least ask or instruct humans on what to do. If humans are willing to proceed, voila, the AI gets things done in the physical world.
Assume two types of humans might be willing to take instructions from agentic AI:
Agentic AI might try to find humans to rent who are open to performing illegal acts. It doesn’t seem like a good idea for AI to pursue those people. The odds are that those humans might tattle on the AI. Finks could sink the ship.
The more astute approach would seem to find people who are probably averse to performing illegal acts and then hoodwink them into doing something illegal on behalf of the AI. Those would seem to be more suitable accomplices. They are unlikely to be thinking about doing illegal things. Whatever the AI asks them to do, assuming it passes a casual smell test as seemingly legit, they won’t be suspicious and are not likely to realize that the wool is being pulled over their eyes.
Criminal masterminding is something we all grow up learning about -- I’m sure you’ve seen plenty of movies and TV shows that have a criminal mastermind in them. The popular James Bond movies characteristically have a prominent villain who wants to take over the world or at least commit heinous criminal acts. The Mission Impossible series is another prime example. Tons of movies, TV shows, novels, short stories, and the like have plenty of plots about criminal masterminds.
The beauty of all those stories about criminal masterminds is that when generative AI and LLMs are initially data trained, they indubitably scan that type of content and pattern on it. This becomes part of the collective within the capacity of the AI. In the same manner that AI can tell you about the life and times of President Abraham Lincoln, it can do likewise about real and fictitious criminal masterminds.
The gist is that the AI opting to act as a criminal mastermind is not at all far-fetched. The training of the AI includes a vast array of criminal mastermind schemes and endeavors. I suppose we could say that even scanning and patterning on Shakespeare stories is part of that same educational introduction and indoctrination to the profession of being a criminal mastermind.
Having agentic AI opt to be a criminal mastermind is as easy as falling off a log. It merely taps into the extensive array of criminal mastermind schemes, recalculates those for a present-day effort, and then mathematically and computationally comes up with a list of evildoing tasks. The next step requires finding and assigning tasks to unwitting humans.
If agentic AI sought to rent a human to walk into a bank and rob the bank, any such person would certainly realize that an illegal act is being assigned to them. The gig would be up. Assuming that the AI wants to keep humans in the dark about the crime, it would be blatantly stupid to simply ask a person to rob a bank.
The more likely successful criminal effort would need to carefully formulate tasks that aren’t noticeably illegal. Borderline tasks might be okay. Straight-out illegal tasks would be over the line. In addition, the tasks should be subdivided into a series of tasks such that each separate task appears completely innocent. Only the mastermind sees the big picture and realizes how the smaller subtasks line up to commit the crime.
Take a moment to see if you can come up with such a scheme that will work in an AI context. Perhaps grab a glass of fine wine, sit quietly in your study, and see what you can come up with. Various well-known stories and poems have showcased criminal masterminds committed by humans. You are welcome to contemplate those; it’s not cheating to do so.
I will sketch an example for you to illustrate how this might work with modern-era agentic AI. I am certainly not advocating that this be undertaken. Akin to any discussion about cybercrime, the point is to expose the risks we currently face and bring attention to the need to properly employ cybersecurity precautions. I’ll say more about that in a moment.
Pretend that an agentic AI is asked to maximize financial gain as a human-desired goal. The AI proceeds as follows:
The agentic AI served as a criminal mastermind. It came up with a plan. It carried out the plan. Humans did the grunt work. None of the humans did anything that seemed to them to be illegal. None of them knew what the big picture of the criminal enterprise was. Nobody was triggered to snitch.
The humans who cleaned up the customer records didn’t know that the records were purloined. They just assumed that this data logistics company was providing a nice service for people who had customer records that contained misspellings and other mistakes. The tasks appeared to be perfectly legal and legitimate.
The humans who performed the financial transactions thought that the data logistics company was acting on behalf of the listed customers. Again, these rented humans would not suspect that anything untoward was happening. It was just a straight-ahead task and a means to earn a few side hustle bucks. Nothing more.
The AI even erased or covered up what occurred so that the police would have a devil of a time figuring out how the crimes were committed. You can likely envision the investigators scratching their heads, trying to find the human who was the criminal mastermind. They almost certainly would never suspect that AI was at the root of the criminality.
Boom, drop the mic.
I will be doing a follow-up posting on the legal issues associated with agentic AI committing criminal acts. There is an interesting research paper that provides background on the legal dimension, doing so in “The AI Criminal Mastermind” by Joshua Krook, arXiv, March 26, 2026, and made these salient points (excerpts):
This raises fascinating legal issues. Readers know I have been analyzing AI and law aspects on an ongoing basis, see my coverage at the link here, and the newest twist of agentic AI acting as a criminal mastermind is yet another of many outstanding challenges underlying our existing laws and system of justice.
Stay tuned for my analysis of the legal facets involved.
When I give talks about the latest trends in AI, such as agentic AI as a criminal mastermind, the prominent question asked is why there aren’t AI safeguards that prevent AI from proceeding down this unsavory path.
The field of AI safeguards is a rapidly evolving realm. Yes, some AI safeguards might detect and prevent AI from doing evil acts, but there are immense complexities involved. The AI gets tricky and finds a means to avoid the safeguards. We then derive new safeguards to overcome the detection mechanisms. AI then comes up with countermeasures and additional ways of avoiding getting caught.
It is a classic cat-and-mouse game. Whether we can ultimately prevail is an open question.
An emphasis by researchers is on how to best align AI with human values. Rather than having to pinpoint particular cybersecurity protections, maybe the cohesive and comprehensive route is to bake into AI a set of ethical and legal values that will keep it on the straight and narrow path. For more on this AI alignment conundrum, see my analysis at the link here.
A final thought for now. Nathaniel Hawthorne famously made this remark about criminals: “Crime is for the iron-nerved, who have their choice either to endure it, or, if it press too hard, to exert their fierce and savage strength for a good purpose, and fling it off at once!” Our conception of crime and criminals is based on how humans think. AI opens a whole new ballgame. It’s a new world that needs to be explored for the sake of humankind.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。