Enterprise AI does not invent new ransomware threats. It accelerates the ones that already exist. That framing is blunter than most vendor communications but it is also more accurate than the typical threat report preamble.
The mechanism is specific. AI assistants and agents deployed across enterprise networks operate with the permissions of the accounts or service principals they run under. If those identities carry excessive access, the AI inherits it. If those identities are compromised, the AI becomes a vehicle for whatever the attacker wants to do next including lateral movement and data exfiltration at a speed no human operator could match. Ransomware operators have always relied on over-privileged accounts to move quickly through a network. A GenAI agent sitting on top of those same accounts removes the manual steps that previously gave defenders time to detect and interrupt an intrusion.
The data leakage problem is separate but compounds the ransomware risk. Check Point reported that 1 in every 54 GenAI prompts sent from enterprise networks posed a high risk of sensitive data exposure and that 15% of all prompts contained potentially sensitive information. Those figures come from Check Point’s own research which means they should be treated as an estimate with a commercial interest behind it rather than a neutral measurement. The directional finding, that employees are feeding sensitive material into AI tools at scale, is consistent with what other researchers have observed. The precise percentages deserve scepticism.
The 91% figure circulating in coverage of this story described as the share of organisations using GenAI tools that are affected by data leakage risk, originates from the same vendor research. I would not build a board presentation around it. What it reflects, even conservatively read, is that data leakage through AI prompts is not an edge-case problem affecting a minority of deployments.
Ransomware Is Rising Independently of AI
The GenAI risk sits on top of a ransomware trend that was already moving in the wrong direction. Industrial Cyber reported a 46% increase in ransomware attacks in the most recent measurement period with education, telecoms and government identified as the hardest-hit sectors. The source for that figure is not a government agency and Industrial Cyber does not specify a precise time window or methodology in the summary available. Treat it as indicative rather than definitive. What is not in dispute is the direction, ransomware volumes have been rising across every credible dataset published in 2024 and 2025 including those from ENISA and NCSC.
The combination matters because ransomware operators do not need AI to run successful campaigns. They need access, time and poorly segmented networks. What GenAI integration does is reduce the time component and increase the access component, two variables that defenders were already losing ground on.
The Permission Problem Is the Core Issue
The specific failure mode worth understanding is not prompt injection or model manipulation which get the most coverage. It is simpler, AI agents are being deployed with the same identity hygiene failures that have plagued service accounts for twenty years. Accounts with standing administrative access, credentials that do not rotate, service principals with permissions scoped to an entire tenant rather than a specific function. Enterprises that have spent years trying to clean up these problems in their human user directories are replicating the same mistakes in their AI deployments, often faster because AI rollouts are happening under pressure and without the same governance scrutiny applied to human access requests.
Least-privilege access is the obvious control and it is mentioned in every advisory on this topic including guidance from Acronis cited by BleepingComputer. The harder question is whether organisations have the visibility to apply it. Most enterprises deploying GenAI tools do not have a complete inventory of what those tools can access which identities they operate under or what data they are handling at the prompt level. You cannot enforce least privilege on a system you have not mapped.
No Named Nordic Incident, But the Exposure Is Real
The source material for this story contains no specific incident involving a named Nordic company and I am not going to generalise from European averages to claim otherwise. What is verifiable is that GenAI adoption among large Nordic enterprises accelerated sharply through 2024 and that the identity governance problems described above are not unique to any geography. Any organisation running Microsoft Copilot, Salesforce Einstein or similar integrated AI assistants on top of an Active Directory environment with legacy permission structures carries this exposure today. The relevant question is not whether Nordic companies are theoretically at risk. It is whether their AI deployments have been reviewed by the same teams that handle identity and access management.
Three Controls That Actually Reduce the Attack Surface
The remediation advice from Acronis via BleepingComputer breaks into three categories that are worth taking seriously even if the source has a product to sell.
- Audit the identities your AI tools run under. Establish what permissions each agent or assistant holds, whether those permissions are scoped to the minimum required and whether the underlying credentials are subject to the same rotation and monitoring policies applied to human accounts. Do this before the next AI deployment, not after.
- Classify what data AI tools can reach. If a GenAI assistant integrated into your document management system can read contracts, personnel files and financial data in a single session, that is not an AI governance problem, it is an access control problem. Data classification and access boundaries need to be defined at the infrastructure level before AI tools are layered on top.
- Build AI usage into your existing security operations monitoring. Prompt-level activity, data accessed during AI sessions and anomalous usage patterns should feed into the same SIEM workflows used to detect compromised user accounts. An AI agent exfiltrating data looks different from a ransomware binary but the access pattern, large volumes of reads across sensitive directories in a short window is recognisable if you are looking for it.
None of these controls require specialised AI security tooling. They require applying disciplines that security teams already understand to infrastructure that many organisations have not yet brought into scope.
References
- How enterprise GenAI can amplify ransomware risk and how to contain it
- Global cyber attacks decline, but ransomware jumps 46% as GenAI threats hit education, telecom, government
- Palo Alto Networks: Top GenAI Security Challenges
This post is also available in:
Svenska
