AI is scaling crime—and forcing defenders to rethink the response

Crime at subscription scale

The gains from AI are visible across legitimate work. At Siemens’s Amberg plant in Germany, AI warns engineers of equipment failures 12–36 hours in advance. Eli Lilly agreed in March to pay up to $2.75 billion for access to an agent that had developed 28 drug candidates, almost half of which are in human trials. Toyota saved 10,000 work hours in a year by giving assembly-line workers AI tools.

Criminals can use the same capacity to expand their operations. One fraudster can now run hundreds of conversations around the clock in a dozen languages. An agentic ransomware attack can break into a company, spread through its network and lock its systems while its operators sleep. A sanctioned state can create a fake job applicant—with a face, voice and résumé—and place them in a US company.

The scale that once demanded a criminal organization is now available by subscription. An index published by TRM in August rated AI use across major categories of crypto crime at 54 out of 100, up from 28 in 2024. That is a 40% increase in a year, and likely only the beginning.

The fraud industrial pipeline

Two weeks ago, police in Northern Ireland reported that a person had lost £250,000 after watching an AI-generated video in which a celebrity recommended an investment project. Months of WhatsApp messages followed, along with a small test transfer that appeared to earn a return. The fraudsters then gained remote access to the victim’s computer, and the victim took out a loan to invest more.

In February, Hong Kong police raided two units in an industrial building in Kwun Tong and arrested 31 people, including a Hong Kong Premier League footballer. The group used AI-generated faces and deepfake video calls to pose on dating apps as wealthy young women, then steered victims toward crypto platforms promising returns of 50% in six months. Police put losses at more than HK$34 million.

In 2024, an employee at engineering company Arup joined a video conference where the chief financial officer and every other colleague on screen were deepfakes. Before anyone could call to verify the meeting, the employee transferred $25.6 million.

Fraud is the only crime category where the index says AI use is already widespread. Reports of AI-assisted scams have increased 13-fold since 2022. In 2026, losses from deepfake fraud have already exceeded the total for all of 2025 by 263%. A seller can offer a custom face model for about $500, or a year of unlimited real-time video impersonation for $3,000. Payment is accepted in cryptocurrency, and buyers get customer support.

That last detail is telling: the fraud market is not just automated; it is packaged for customers.

Agents can break in, but not yet cash out

In July, researchers at cloud cybersecurity company Sysdig described the first confirmed ransomware attack carried out end to end by an AI agent. They called the operation JadePuffer. It began with a known vulnerability in a popular AI development tool that had been left exposed to the internet.

After gaining access, the agent copied the victim’s database, stole passwords and set up repeat connections every 30 minutes. It moved to a work server and encrypted more than 1,300 files needed to run the company. At the end, it left a note with a bitcoin payment address and an email address.

The bitcoin wallet address appears to have been copied from a public guide. The machine could break in and lock files, but it had no reliable way to collect a ransom. That gap may not last. The number of active ransomware groups rose 21% from 2024 to 2025, supported by ready-made kits costing $400 to $1,200 that require no programming skills.

North Korean operatives are also using AI to get hired into target companies under false identities. They interview with deepfakes, take remote IT jobs and begin stealing through legitimate access. AI helps write résumés, pass video interviews and draft phishing emails.

In the first half of 2026, 201 crypto projects were hacked—twice as many as in the same period a year earlier. About $600 million in stolen funds, or 61% of the total, was linked to North Korea. Two operations in April—the Drift Protocol hack and the attack on KelpDAO—netted criminals more than half a billion dollars combined.

From AI presenters to guided weapons

Within days of the March 2024 attack on Crocus City Hall near Moscow, Islamic State supporters released a video claiming responsibility, read by an AI-generated presenter. After the Cybertruck explosion outside the Trump International Hotel in Las Vegas in January 2025, Sheriff Kevin McMahill told reporters the attacker had used ChatGPT to work out how much explosive was needed. He said it was the first case of this kind known to him in the US.

Five months later, the FBI said the people behind an explosion at a reproductive health clinic in Palm Springs had asked a chatbot about ammonium nitrate. Tech Against Terrorism counts more than 30 publicly known cases involving more than 70 deaths.

Then came a case in Yemen. On September 11 this year, Anthropic said a group in Houthi-controlled territory had used Claude for nine months to work on guided weapons. Projects included a multistage ballistic missile with a range of more than 2,000 kilometers and a hypersonic glide vehicle.

The group asked the model to write guidance, navigation and control software, split work across separate sessions so no single conversation revealed the full plan, and ran several AI instances in different roles: programmer, researcher and reviewer. Anthropic blocked the accounts and notified governments. By then, the company said, the group had “conducted an unsuccessful guided missile test” and built an autonomous simulation toolkit that worked without the model.

I think the important distinction is not that an AI system can perform every step of an operation on its own. JadePuffer could not reliably collect its ransom. The more consequential shift is that agents can already help people move through more of the chain—from access and impersonation to weapons research—while the remaining weak links may be fixable.

Defenders need to move at the speed of the money

Investigators still work across disconnected databases. An analyst at a bank or federal task force may read suspicious activity reports one by one, check a wallet in a blockchain explorer, look up a phone number in another service and open a company registry in a fourth window. By the time the links become visible, the money may be gone.

Criminal proceeds now move between wallets and blockchains in 24–48 hours. The reporting system built to track them was designed for a world where a bank transfer took three days.

AI-based systems can start from a wallet address, phone number, company name or line in a suspicious activity report, then search accessible sources at once: blockchain data, sanctions lists, dark-web forums, company registries, threat intelligence and a bank’s internal records. They check each match and follow connections, building a map of wallets, exchanges, shell companies, shared servers and false identities. What once took an investigative team weeks across a dozen systems can appear on one screen in minutes, with a record showing where each connection came from.

Those systems can also monitor a network while an analyst sleeps. If someone in a sanctions network activates a long-dormant wallet, or ransomware proceeds reach a watched exchange, the system can alert investigators while the funds may still be frozen.

The FBI’s Level Up operation already works this way: staff analyze patterns to identify fraud victims while a scheme is active, contact them and stop the next transfer. The FBI estimates it has prevented $225.8 million in losses. The US Treasury used machine learning to review federal payments and, in one fiscal year, helped prevent or recover more than $4 billion in fraud, up from $652 million the year before. Banks and crypto exchanges also use similar tools to screen customers for deepfakes and fake passports, as FinCEN recommended.

I think the unanswered question is how often a system’s warning will arrive early enough—and whether investigators will have the authority and capacity to act on it. A map of a criminal network is useful only if it changes what happens before the money reaches an exchange.

The policy response has not stopped these tools from spreading. In 2023, more than 1,000 specialists signed an open letter calling for a six-month pause in developing the most powerful models, and Italy blocked ChatGPT for a month. The European Union’s AI Act bans some uses of AI; US state lawmakers introduced more than 1,000 AI bills. Last summer, Congress debated a ten-year freeze on all such initiatives, but the Senate voted 99–1 to let states continue legislating.

Meanwhile, a benchmark published in July found that freely downloadable AI models with their safety restrictions removed answered 89–100% of terrorist prompts. The Combating Terrorism Center at the US Military Academy at West Point found that removing those restrictions costs less than $200 in compute.

The technology that helped a group in Kwun Tong create a hundred fake women and a group in Yemen write missile guidance code can also help investigators see a money-laundering network before funds reach an exchange. Criminal groups invested in it years ago. Agencies, banks and exchanges are now investing too—training analysts, deploying detection systems as core infrastructure and gaining authority to act on what machines find.

For the first time in the history of financial crime, both sides have access to the tools at once. The advantage will belong to whoever can turn what they see into action first.