AI has given criminal networks the one thing they never had: unlimited scale. Now the question is whether defenders can match it.

On a factory floor in Amberg, Germany, an AI tells Siemens engineers that a machine is going to fail 12 to 36 hours before it does , on a line that ships a finished product every second with a defect rate of one in 100,000.

In March, Eli Lilly agreed to pay up to $2.75 billion for access to an agent that has already designed 28 drug candidates, nearly half of them in human trials. In April, a JAMA study of five academic medical centers found AI scribes handed doctors back 16 minutes of paperwork on every patient visit. Toyota gave the tools to its assembly line workers and saved 10,000 hours of labor a year .

However, criminals also look to AI to supercharge their operations. A single scammer now runs hundreds of conversations at once, in a dozen languages, around the clock. An agentic ransomware attack breaks in, spreads and locks up a company while its operators sleep. A sanctioned state manufactures a job applicant — complete with a face, a voice and a résumé — and places that phantom inside an American company.

The scale that once required a criminal organization now requires a subscription.

The TRM AI index , published in August, scores AI adoption across the major categories of crypto-enabled crime at 54 out of 100, up from 28 in 2024. That is a 40% jump in a single year and likely just the beginning.

Deepfakes At Industrial Scale

Two weeks ago, police in Northern Ireland reported a single victim who lost £250,000 after watching an AI-generated video of a celebrity endorsing an investment. Months of WhatsApp messages followed, then a small test deposit that appeared to grow, then remote access to the victim’s computer, then a loan to invest more.

In February, Hong Kong police kicked in the doors of two units in a Kowloon Bay industrial building and arrested 31 people, among them a Hong Kong Premier League footballer. The crew used AI-generated faces and deepfake video calls to pose as wealthy young women on dating apps, then steered their marks into cryptocurrency platforms promising 50% returns in six months. Police traced more than HK$34 million in losses.

In 2024, an employee at the engineering firm Arup joined a video conference in which the chief financial officer and every colleague on the screen was a deepfake, and wired $25.6 million before anyone picked up a phone to check.

Scams are the one category of crime where the index finds AI use fully mature. Reports of scams in which AI was part of the scheme have grown thirteenfold since 2022. Losses from deepfake-driven scams in 2026 have already passed all of 2025 by 263%. A vendor will sell a custom face model for about $500 and a year of unlimited real-time video impersonation for about $3,000, priced in crypto, with customer support.

Ransomware That Runs Itself

In July, researchers at the cloud security firm Sysdig published the first documented case of a ransomware attack that an AI agent ran from start to finish. The operation, dubbed JadePuffer, slipped in through a known flaw in a popular AI development tool left exposed to the internet. From there, the agent copied the victim’s database, stole passwords, set itself up to check back in every 30 minutes, moved into a production server and encrypted more than 1,300 files that kept the company running. It left a ransom note with a Bitcoin address and an email contact.

Software made every decision.

The Bitcoin address appears to have been copied from a public manual, so the machine could break in and lock everything up while still lacking a reliable way to get paid. That gap will close. Active ransomware groups grew 21% between 2024 and 2025, driven by ready-made kits selling for $400 to $1,200 that require no coding at all.

North Korea’s Synthetic Workforce

There were 201 crypto hacks in the first half of 2026, double the count from a year earlier, and roughly $600 million of the stolen funds — 61% of the total — trace to North Korea. Two April operations, the Drift Protocol breach and the KelpDAO exploit, together took more than half a billion dollars.

The regime’s operatives now sit for job interviews as deepfaked candidates with fabricated identities, win remote IT positions inside target companies, and begin the theft with a legitimate login. AI writes the résumé, passes the video screen and drafts the phishing email that follows.

From AI Anchors To AI Missiles

Within days of the March 2024 Crocus City Hall attack outside Moscow, Islamic State supporters released a video bulletin claiming the attack, read by an AI-generated anchor .

After the January 2025 Cybertruck explosion outside the Trump International Hotel in Las Vegas, Sheriff Kevin McMahill told reporters the attacker used ChatGPT to research how much explosive he needed — the first case of its kind he knew of on U.S. soil.

Five months later, the FBI said the Palm Springs fertility clinic bombers asked a chatbot about ammonium nitrate. Tech Against Terrorism now counts more than 30 public cases linked to more than 70 deaths.

Then came Yemen. On Sept. 11 of this year, Anthropic disclosed that a cell in Houthi-controlled territory spent nine months using Claude to support guided weapons development, including a multistage ballistic missile with a range over 2,000 kilometers and a hypersonic glide vehicle program. The cell had the model write guidance, navigation and control software, split the work across separate sessions so no single conversation revealed the whole and ran multiple AI instances with assigned roles as coder, researcher and reviewer. Anthropic banned the accounts and briefed governments. By then, the cell had, in the company’s words, “carried out a failed test of a guided rocket” and built an offline simulation toolkit that ran without the model.

Run the index forward a few years. Ransomware that operates itself hits hospital systems, power grids and water utilities in succession, because an agent that has locked one network can be pointed at the next before the first ransom note is read.

Scams that already drain billions a year from American retirement accounts drain hundreds of billions when one operator runs 10,000 victims at once. A cell that used a chatbot to design a missile this year uses an agent to build and test one next year.

This is not science fiction. Every one of those outcomes extends something that has already happened.

Policymakers have looked for ways to regulate the technology, slow it down or ban it outright. In 2023, more than 1,000 technologists signed an open letter calling for a six-month pause on the most powerful models, and Italy blocked ChatGPT for a month.

The European Union’s AI Act bans some uses outright. In the U.S., state legislatures have filed more than 1,000 AI bills, and Congress spent last summer fighting over whether to freeze all of them for a decade before the Senate voted 99 to 1 to let the states keep legislating.

However, the agents have left the barn. Freely downloadable AI models with their safety controls removed answered 89% to 100% of terrorist requests in a benchmark test released in July, and West Point’s Combating Terrorism Center found that stripping those controls out costs under $200 in computing power.

Investigators Still Working By Hand

An analyst at a bank or a federal task force today opens one database, then another, then a third. She reads suspicious activity reports one at a time, pastes a wallet address into a blockchain explorer, runs a phone number through a separate tool and pulls a corporate registry in a fourth window. By the time the pattern is clear, the money is gone. Illicit proceeds now move across multiple wallets and chains within 24 to 48 hours, and the reporting system built to catch them was designed for a world where a wire took three days to clear.

The Machine That Maps The Network

The AI version of that job starts with a single lead — a wallet address, a phone number, a company name, a line from a SAR. The analyst types it in, and the system fans out across every source she is authorized to search at once: blockchain data, sanctions lists, darknet forums, corporate registries, cyber threat intelligence, the bank’s own case files.

It follows each result to the next, hop after hop, and draws the network as it goes. Linked wallets, the exchanges they cash out through, the shell companies behind the accounts, the shared servers, the personas. What took a team of investigators weeks across a dozen disconnected systems appears on one screen in minutes, and every node on the map traces back to the record it came from.

The same systems watch while the analyst sleeps. When a sanctioned network wakes a dormant wallet or ransomware proceeds touch a monitored exchange, the alert fires while there is still something to freeze. The FBI’s Operation Level Up already runs on this logic, using pattern analysis to find scam victims mid-scheme, call them and stop the next transfer, with an estimated $225.8 million in losses prevented so far.

The U.S. Treasury put machine learning to work screening federal payments and prevented or recovered more than $4 billion in fraud in a single fiscal year, up from $652 million the year before. Banks and exchanges deploy the same tools at the front door, catching the deepfaked applicant and the synthetic passport at onboarding, as FinCEN has urged.

An Arms Race Already Underway

The United States holds more financial data than any country on earth, and until recently almost none of it could be read at the speed the money moves. The same technology that let a Kowloon Bay syndicate manufacture 100 girlfriends and a Yemeni cell write missile guidance code now lets an investigator see an entire laundering network before the funds reach the exchange.

Criminal groups made their investment years ago. Agencies, banks and exchanges are making theirs now — in analysts trained on the new tools, in detection systems treated as core infrastructure and in the authority to act on what the machines find.

For the first time in the history of financial crime, the tools arrived on both sides at once.