AI Lowers the Ransomware Barrier: One Attacker, No Expertise, Full Campaign

Generative AI is quietly dismantling one of the last practical barriers that kept ransomware in the hands of organized crews. According to cybersecurity researchers whose findings were compiled by Axios and the threat lab at SentinelOne, large language models now write exploit code, translate stolen data, draft ransom notes and even handle negotiations with victims — allowing a single individual with minimal technical background to run an operation that once demanded a coordinated team.

  • Speed over sophistication: In one documented case, an AI agent rewrote exploit code in 31 seconds; in another, a lone attacker compressed a campaign that normally takes weeks into 72 hours.
  • Acceleration, not revolution: SentinelOne’s researchers argue the models are speeding up the ransomware cycle rather than transforming it from the ground up.
  • The real risk is volume: As the skill floor drops, the number of would-be attackers rises — and defending against a thousand clumsy amateurs can be as draining as fending off ten professionals.

What AI actually changes in the attack chain

The core capability shift is about who can execute, not what is technically possible. A senior cybersecurity specialist at the firm Sysdig put it bluntly: someone with no ransomware skills at all could plausibly carry out this same type of operation. In the case her company detected, a single attacker used AI agents to automate most of the intrusion — writing the exploit, exfiltrating data and negotiating the ransom — with the agent rewriting code in 31 seconds. A separate investigation by the firm Sygnia found a solitary attacker who squeezed a campaign that would normally take weeks into 72 hours.

SentinelOne’s analysis highlights where the models add the most leverage: working across more languages and systems, flagging sensitive files in a language the attacker cannot even read, producing more convincing phishing lures, and evading security controls by spreading requests across multiple sessions. This is not speculative fiction — a comparable jump was already observed when an autonomous AI agent struck Adif and Renfe and made off with 500 GB of data.

Amateur hour, with real consequences

The picture is messier than the superhacker narrative suggests. Many of these AI-assisted attacks remain sloppy: exposed servers, poorly disguised activity and, in one instance, a bitcoin address possibly invented by the model itself. SentinelOne’s researchers stress that every step in the documented cases was already achievable by well-resourced groups — what has changed is how few people are needed to attempt it.

Additional reporting on the emerging group tracked as N0n reinforces the paradox. The crew spent more than $7,000 on generative AI tools and leaned on OpenCode, an open-source platform for working with different models, to build reconnaissance programs. Yet when one of its servers was accidentally exposed, investigators found the group had listed just over 20 actual targets against the hundred-plus victims it advertised on its leak site. Researchers also found an encryption program but could not confirm it was used systematically, and the group’s own security failures — including reused encryption keys — undercut its claims.

Someone with no ransomware skills at all could plausibly carry out this same type of operation.

The defensive window is narrowing

The more consequential danger, according to SentinelOne, lies elsewhere: attackers migrating to open-source models running on their own servers, beyond the safety guardrails of commercial providers. That would leave defenders with far less visibility into what tools are being used and how. It is the dark side of a single coin — the same capabilities offered for defense, such as OpenAI’s GPT-5.4-Cyber release for thousands of security defenders, work just as well for offense. And the targets remain familiar: healthcare, critical infrastructure and personal data, as demonstrated by the theft of medical records and fingerprints belonging to 1.8 million people at the largest public health system in the United States.

The sensible reading avoids both extremes. The myth of the all-powerful AI superhacker is, for now, overstated — real cases are full of beginner mistakes, and those selling panic are often selling the remedy too. But the metric that matters is not sophistication; it is headcount. When one unskilled person can attempt what previously required a team, attack volume climbs sharply. Defenders still have time, but that window is closing at the same pace AI becomes easier to use. The researchers’ own advice is measured rather than alarmist: don’t panic, but start preparing now, in small steps. The barrier that once kept amateurs out has just come down, and it is not going back up.

Sources & References