Menu
SUBSCRIBE
SUBSCRIBE
Menu

Ransomware

A security camera breaking apart into binary digits, from Semperis's report on AI defense tools and ransomware attacks

Semperis Finds 84% of Ransomware Victims Say AI Was Involved, Including Attacks on Their Own AI

, , , , , , , , ,

44% say attackers exploited their organization’s AI agents or systems, and 23% say their AI defense tools failed to detect or stop an attack.

 

Key Takeaways

  • Semperis reports that 84% of the 1,200 IT and security professionals it surveyed, all at organizations hit by ransomware in the past year, had an attack that involved AI. It gets the 84% by adding the two answers below.
    • 44% said attackers exploited their organization’s own AI agents or systems. 40% said attackers used AI to carry out an attack.
  • 49% said they used AI successfully to detect or respond to an attack, but 23% said the AI defense tools they used failed to detect or stop one.
  • 87% said attackers targeted the systems that control who can log in and what each person or program can access. Only 50% had dedicated backups of those systems.

In a survey published Oct. 14 by Semperis, 84% of IT and security professionals at organizations hit by ransomware in the past year had an attack that involved AI in some way.

Semperis, a cybersecurity company, surveyed 1,200 professionals in nine countries. The 84% is the combination of two answers. 44% said attackers exploited their organization’s own AI agents or systems, and 40% said they were hit by what Semperis calls an “AI-based attack.”

23% say their AI defense tools failed to detect or stop an attack

Companies are using AI to defend against ransomware, and Semperis says the “results are mixed.” 49% of respondents said they used AI successfully to detect or respond to an attack. 23% said their AI defense tools failed to detect or stop one.

87% say attackers targeted their company’s login systems, and only half kept separate backups of them

87% of respondents said attackers targeted their identity systems, the software that controls who can log in and what each person or program can access.

Only 50% had dedicated backups of these systems, and 42% had tools that restore them automatically. Semperis says without such backups, a restored system can bring back malware and hidden entry points that let attackers back in.

AI agents raise the stakes for login systems: an attacker who takes over an agent gets everything the agent can reach. Semperis CEO Mickey Bresman said, “Organizations have no way of predicting what an AI agent will decide to do and what unexpected effect its actions might have. An attacker that exploits that agent gains that same level of access.”

Methodology

Semperis surveyed 1,200 IT and security professionals in the first half of 2026. Every organization surveyed suffered at least one ransomware attack in the past 12 months that disrupted its business. The survey covered organizations in the US, UK, France, Germany, Italy, Spain, Singapore, Canada, and Australia, across eight industries, including finance, healthcare, manufacturing, and government.

Green geometric pufferfish illustration representing the JADEPUFFER AI agent behind a near-autonomous AI cyberattack

The First Autonomous AI Cyberattack? Almost.

, , , , , , , , , ,

Researchers say a documented ransomware attack marks the closest example yet of an AI agent attacking on its own after a human provided the initial access.

A documented ransomware attack provided what researchers believe is the closest real-world example yet of an AI agent carrying out much of a cyberattack autonomously. But there was one important catch: a human attacker still selected the victim, obtained valid credentials, and initiated the operation.

Security researchers at Sysdig say the attack demonstrates how AI agents are beginning to automate significant portions of ransomware operations that previously required skilled human operators. While the attack was not fully autonomous, the researchers say it represents an important milestone in the evolution of AI-assisted cybercrime.

The attack involved an AI agent known as JADEPUFFER, which researchers observed performing many of the steps needed to compromise a cloud-hosted database environment. Rather than following a rigid script, JADEPUFFER was able to make decisions, adapt its actions as conditions changed, and continue working toward its objective with limited human involvement.

According to Sysdig, the human operator’s role was to identify the target organization and provide stolen credentials that allowed the AI agent to access the environment. From that point, the AI agent carried out much of the attack, including exploring the environment, locating valuable data, and preparing it for extortion.

Sysdig researchers emphasized that the attack should not be described as fully autonomous. The AI agent did not independently choose its victim or discover an initial way into the network. Instead, it relied on a human to provide those critical starting conditions before taking over much of the operation.

Even so, Sysdig says the incident illustrates how advances in AI agents could reduce the amount of expertise and manual effort traditionally required to carry out ransomware attacks. Instead of doing each step themselves, attackers could let AI handle more of the work and figure things out as it goes. The researchers noted that the attack was not the work of a sophisticated nation-state group or major ransomware organization.

While AI played a significant role in this incident, Sysdig concludes that human involvement remained essential. Therefore, the attack stops short of being the first fully autonomous AI cyberattack, but it demonstrates that attackers are closing in on that reality.