The pitch writes itself: artificial intelligence will let criminals run ransomware at machine speed, and the internet will drown. That prediction is now partly fulfilling itself — not through autonomous malware, but through something far more mundane and, in a strange way, more instructive. The clearest gains from AI in this story belong to the attackers, and they are gains in persuasion and scale, not in code. Understanding exactly where the lift happens is the first step toward deciding what can be fixed and what cannot.
The Numbers Behind the Claim
Ransomware and data extortion have ranked among the most damaging cybersecurity threats to institutions over the past decade, and the past few months have pushed the volume higher. Researchers who track and attempt to de-duplicate leak-site claims broadly agree that alleged incidents have spiked since spring. Some put the count of claimed attacks posted on leak sites at nearly 3,000 over the summer, which by some estimates amounts to a 20 to 40 percent increase. [1] Volume always fluctuates, but the researchers emphasize that the overall level has remained consistently and concerningly elevated for years. The recent jump is therefore a change in degree on top of an already high baseline, not a new phenomenon.
Cynthia Kaiser, a former FBI cyber deputy director who now leads research for the anti-ransomware platform Halcyon, draws a line between the domestic and international picture. [1] The United States has stayed comparatively steady, while attacks abroad have climbed sharply — and that international surge is, in her assessment, the genuinely new element. She attributes part of it to AI. The geography matters: a criminal who once needed local language skills to run a convincing campaign can now operate across borders with content that reads as native.
The most important qualification comes from Allan Liska, a longtime ransomware researcher and threat intelligence analyst at Recorded Future, who stresses that the successful groups remain largely human operated. Fully AI-powered ransomware attacks have yet to emerge, he notes. The criminal business model already works well enough that there is little pressure to automate it end to end. What has changed is the experimentation: groups are testing AI heavily across the attack chain, which means the picture could shift, and shift quickly.
What the Models Actually Do
The concrete gains sit in the parts of the job that were always laborious. Liska points to better phishing emails and the negotiation process as areas where AI is relied on more heavily. [1] Both are language tasks. A phishing message has to clear a low but real bar of plausibility; a negotiation has to sound like a counterparty with authority and patience. These are exactly the tasks a language model performs cheaply and at volume, and exactly the tasks that previously consumed human hours.
Kaiser connects ransomware to the broader scamming economy, because initial access often comes through the same phishing and social engineering techniques. Japan offers the sharpest illustration. Successful scamming there has spiked drastically in recent months as AI enables foreign attackers to generate consistently convincing content in Japanese for the first time. That is the lift in its purest form: a language barrier that once functioned as a natural defense has been lowered, and the attackers did not need to learn the language to cross it.
The Teenager and the $800,000
The most striking consequence of that lowered barrier is organizational. Behind the established names — the Gentlemen, Qilin, and Akira have ranked among the most prominent and stable operations in recent months — sits a widening field of smaller groups and new entrants. Halcyon’s analysis counts around 111 ransomware groups making claims between July and September of this year. One newcomer calling itself Global Secret Group claimed 48 victims, more than half of them listed on the same day, according to analysis by the security firm ReliaQuest. Another, Storm, listed 71 victims in 20 days. Not every claim is necessarily legitimate, but even inflated claims create uncertainty for the businesses named.
Brandon Tirado, director of threat research at ReliaQuest, sees the acceleration as the engine of that fragmentation. AI, in his view, has allowed threat actors to go out and start their own groups rather than remain beholden to traditional ransomware-as-a-service providers. His team recently identified the operator behind one new group and found an inexperienced teenager who had made more than $800,000 in under two weeks. He declined to name the teenager publicly. Whatever else that figure says, it says the barrier to entry has fallen far enough that a novice can clear it.
Kaiser agrees that more actors are using AI to strike out on their own, with expanded capabilities to trick targets, exploit vulnerabilities, check their code, and even compromise other groups’ remote-access tools to gain redundant access in case one avenue gets blocked. She is careful about the long run. Many of these upstarts may surge and then burn out, she suspects, because they lack the institutional knowledge and resources of an established group. The lift is real; durability is a separate question.
The Simplification Nobody Expected
One of the more counterintuitive effects concerns method. Established attackers typically run double extortion: they deploy traditional ransomware that encrypts a victim’s systems while also exfiltrating data and threatening to publish it. Independent operators, by contrast, may lean toward data extortion alone, because it simplifies their operations. The encryption step — building, deploying, and managing working ransomware — is the technically demanding part. Theft and threat are not.
Liska’s observation captures the shift precisely: the number of attacks Recorded Future has seen this year already dwarfs the total for 2025, yet the number of attacks involving encryption may actually decrease year over year. The number of attacks Recorded Future has seen this year already dwarfs the total for 2025, yet the number of attacks involving encryption may actually decrease year over year, he says. More incidents, less encryption. That is not the shape of an AI that has learned to write better malware. It is the shape of an AI that has made the human-facing parts of extortion — the email, the pressure, the negotiation — cheap enough that some operators skip the hard technical work entirely.
The Barrier That Stays Standing
Here is the part that resists a technical fix. Every gain described above is a gain in language and coordination, and language and coordination are precisely what may not be easily patched out of a model. You can harden a server; you cannot harden the fact that a convincing message in Japanese is now cheap to produce. The defense that would matter most is institutional: victims reporting incidents, researchers de-duplicating claims, and organizations sharing what they learn. Yet victims do not always disclose, and attackers inflate their numbers, which is why exact counts remain impossible to establish.
That is the ethical knot no model resolves on its own. The same language capability that lets a foreign attacker write fluent Japanese could equally let a defender write a clear warning to a vulnerable population. The difference is not in the tool. It sits in who wields it and what the surrounding institutions permit. The prediction that AI would supercharge ransomware is, in this reading, coming true, but not in the way the headline implied: the machine did not take over the crime. It took over the words, and the words were the part that always required a human.