Ransomware Victims Say AI Showed Up in 84% of Attacks

The most quoted number from Semperis's 2026 Ransomware Risk Report is 84%. That is the share of ransomware attacks described by surveyed victims in which AI was weaponized in some way. The report was released on October 14, 2026, and the figure travelled fast.

It deserves a closer read than it got. The headline captures something real about how ransomware has changed. But the details point to a less cinematic and more useful conclusion: AI is amplifying attacks that still mostly succeed through old weaknesses.

The survey

Semperis, an identity security company, polled about 1,200 IT and security professionals across nine countries and eight industry sectors. It is a vendor-sponsored survey of reported experiences, not forensic attribution of individual campaigns, and the numbers should be read that way.

The key findings, as reported:

• AI was weaponized in some way in 84% of the ransomware attacks respondents described.
• 44% of victims said attackers exploited their own AI agents or systems.
• 87% of attacks compromised identity infrastructure.
• More than 69% of victims experienced multiple attacks.
• About 59% of those hit said they had paid a ransom at least once.
• More than three-quarters of victimized employees reportedly experienced physical threats or psychological pressure.

Two kinds of AI involvement

The 84% figure bundles two very different things, and pulling them apart is where the report gets interesting.

The first is attacker-side AI: using models to scan networks for weaknesses faster, write convincing phishing messages, and tailor social engineering to specific employees. A Semperis release for Singapore put attacker-side AI use at about 40% of attacks globally. This is the version most headlines imagined, and it is significant, but it is a minority of cases.

The second is the victim's own AI being turned against them. In 44% of cases, respondents said attackers exploited the organization's AI agents or systems. That is the newer and arguably more alarming finding. Companies have spent the past two years wiring AI agents into email, ticketing, file storage and internal tools. Each of those agents holds credentials and permissions. To an attacker, an over-privileged agent is just another account to take over, and often one that nobody is watching closely.

Semperis researcher Tomer Bar has also warned that AI could help ransomware developers write more sophisticated malicious code, speed up development and complicate investigations. That is a forward-looking concern rather than a measured result in this survey, but it fits the pattern: AI lowers the skill floor for attackers and raises the speed ceiling for the capable ones.

Identity is still the front door

The number that should shape security budgets is not 84%. It is 87%, the share of attacks that compromised identity infrastructure: directory services, single sign-on, privileged accounts.

That is consistent with how breaches have worked for years. Attackers get in with stolen or weak credentials and escalate through misconfigured identity systems. AI makes reconnaissance faster and phishing more persuasive, but it is largely accelerating a familiar playbook rather than replacing it.

It also connects the two halves of the AI finding. An AI agent is an identity. If it has broad permissions and weak monitoring, compromising it is an identity attack with a new label.

The human cost

The finding that more than three-quarters of victimized employees faced physical threats or psychological pressure deserves more attention than it received. Ransomware crews have increasingly targeted individuals directly, contacting staff and executives to push for payment. AI-generated voice and personalized messages make that pressure cheaper to apply at scale.

The payment rate is the other sobering figure. With roughly 59% of victims admitting they had paid at least once, and more than 69% suffering repeat attacks, the economics still favor the attackers. Paying does not guarantee data returns or that the victim will be left alone, as the report notes.

What it means for teams deploying agents

The practical takeaways are not new, but the report gives them fresh weight:

• Treat agents as privileged identities. Give them the narrowest permissions possible, rotate their credentials, and log what they do.
• Harden identity infrastructure first. If 87% of attacks run through it, it is the highest-leverage place to spend.
• Assume phishing quality is now uniformly high. Training employees to spot bad grammar is obsolete.
• Plan for repeat attacks. Most victims are hit more than once.

The read

The easy story is that AI has made ransomware unstoppable. The data supports a narrower and more actionable one. Attackers are using AI to work faster, and they are increasingly exploiting the AI systems companies have rushed to deploy. But they are still mostly getting in through identity, the same door security teams have been told to lock for a decade. The new risk is that every AI agent a company adds is another key to that door.