New RAID-AI Framework Empowers Cybersecurity Leaders to Defend Against AI-Targeted Ransomware

As artificial intelligence becomes deeply embedded in the operational fabric of modern enterprises, the threat landscape surrounding these complex systems has shifted dramatically. Cybersecurity leaders now face a dual challenge: protecting their standard IT infrastructure while simultaneously shielding sensitive AI models and datasets from increasingly sophisticated extortion attempts. To address this mounting concern, a new, free-to-use framework has been launched to help organizations identify and mitigate the specific vulnerabilities that could leave their AI systems exposed to ransomware.

The tool, known as the Ransomware Assessment and Incident Defense Framework for AI Systems (RAID-AI), is now fully available online, providing security teams with a structured approach to evaluate their defensive readiness. The project is the culmination of two years of intensive development led by Behnaz Karimi and Yuvaraj Govindarajulu, both of whom serve as co-leads of the Open Worldwide Application Security Project’s (OWASP) AI Bill of Materials (AIBOM) initiative.

The framework was originally slated for a formal introduction during the ISC2 Security Congress held in Denver, Colorado, in late October. However, due to unforeseen visa complications, Karimi was unable to attend the event in person. Despite this hurdle, the release of the framework has proceeded, offering a vital resource to a global community of cybersecurity professionals grappling with the unique intersections of AI security and ransomware prevention.

A Manual Approach to Ransomware Readiness in AI

In an interview with Infosecurity, Karimi clarified the core philosophy behind RAID-AI. The framework is designed to address ransomware threats targeting AI systems regardless of whether the attackers themselves are utilizing AI tools to facilitate the breach. Unlike automated scanning tools that attempt to connect to a company’s network or probe for weaknesses via external probes, RAID-AI takes a manual, assessment-based approach. It requires security teams to conduct a rigorous self-examination of their existing controls, fostering an internal culture of security awareness and accountability.

The urgency for such a tool is underscored by the evolving nature of cyber-attacks. "Right now, we are slowly seeing that ransomware attacks targeting AI systems are happening," Karimi noted. As AI models become critical intellectual property and core components of decision-making engines, they are increasingly becoming prime targets for threat actors.

Karimi pointed to several troubling developments that highlight this trend. She referenced an April 2025 report authored by Zoran Mitrovic, head of engineering at Kintra Health, which detailed how malicious actors are employing sophisticated data poisoning techniques to compromise the integrity of AI systems, effectively holding them hostage or rendering them useless. Furthermore, the industry is still reeling from the implications of an alleged cyber-attack in June 2026 targeting Novo Nordisk. In that incident, hackers claimed to have successfully exfiltrated proprietary AI models and manufacturing recipes, highlighting the catastrophic potential of intellectual property theft when AI assets are not adequately protected.

The RAID-AI assessment is comprehensive, covering both general cybersecurity best practices and the nuances of AI-specific risks. It focuses on the protection of AI models, the integrity of the data used for training and inference, and the security of the supporting infrastructure that allows these systems to function.

The framework is delivered through an accessible web console designed to cater to a diverse group of stakeholders, including Chief Information Security Officers (CISOs), cybersecurity officers, and hands-on security engineers. The assessment process is granular; users manually evaluate their security controls by selecting maturity levels on a scale ranging from zero to five. These controls are categorized into both general and AI-specific domains, with defined maturity tiers labeled as basic, advanced, and high, allowing organizations to visualize their security posture relative to industry standards.

General controls within the framework encompass essential pillars such as incident recovery, employee training, and the protection of sensitive data. In contrast, the AI-specific categories delve into the technical complexities of data handling, the security of machine learning tools, and the protection of models during runtime. By integrating these two perspectives, RAID-AI ensures that organizations do not neglect fundamental security hygiene while simultaneously addressing the novel threats introduced by AI deployment.

A significant strength of the RAID-AI framework is its interoperability with established security standards. The developers have mapped the framework’s controls to two prominent voluntary security frameworks: the US National Institute of Standards and Technology’s (NIST) Cybersecurity Framework 2.0 (CSF 2.0) and the US Cybersecurity and Infrastructure Security Agency’s (CISA) Cybersecurity Performance Goals 2.0 (CPG 2.0). By aligning with these frameworks, RAID-AI allows organizations to integrate their AI-specific assessments into broader organizational security reporting.

Furthermore, the tool provides a bridge to the adversarial perspective. It allows users to map their identified risks against three established attack-mapping frameworks: MITRE ATT&CK, MITRE ATLAS, and the NIST Adversarial framework. This mapping capability helps security teams understand how their specific AI vulnerabilities might be exploited by real-world threat actors, providing a clearer picture of the threat landscape. For those looking to implement the tool, Karimi and Govindarajulu have provided comprehensive documentation, which includes detailed guidelines on usage, the methodology behind the development of the framework, and clear explanations of how maturity levels and security scores are calculated.

Future Plans: Expanding into Agentic AI Systems

Because the RAID-AI assessment does not require organizations to connect their internal systems to the web console or upload proprietary documents, it avoids the security risks associated with sharing sensitive data with third-party platforms. It is not an automated audit tool; instead, it serves as a structured, guided, and manual self-assessment. Because the accuracy of the output is entirely dependent on the rigor and honesty of the users’ self-evaluations, the tool does not provide an independent verification of an organization’s security posture.

Instead, the framework is designed to be a catalyst for internal collaboration. Karimi emphasized that the most effective use of the tool involves bringing together diverse teams—including security managers, system engineers, and AI-security specialists—to participate in workshops. By collectively reviewing their current security practices and documenting them within the console, these teams can foster a more cohesive understanding of their risks.

The console provides a centralized dashboard where results can be visualized, allowing for easy identification of gaps in security maturity. The platform also enables users to generate and download PDF reports that highlight identified weaknesses, which serves as a valuable resource for leadership when prioritizing security investments. While the dashboard provides senior executives with a high-level view of the organization’s security health, the underlying technical data serves as a roadmap for security engineers to guide deeper, more granular discussions about technical improvements.

Looking ahead, the project’s scope is set to grow. Karimi expressed a desire to provide more practical, actionable resources for users, including instructional videos and security blueprints. These resources are intended to help security engineers move beyond assessment and into implementation, providing them with the necessary context to apply individual controls effectively.

Perhaps most significantly, the developers are already looking toward the next frontier of AI development. As organizations begin to deploy agentic AI—systems capable of performing tasks and making decisions autonomously with minimal human intervention—the nature of ransomware threats is expected to evolve in tandem. Karimi and her team are currently in the process of developing a version of the tool specifically tailored to address the unique ransomware risks posed by these agentic AI systems.

As the digital landscape continues to incorporate more autonomous intelligence, the work of initiatives like RAID-AI becomes increasingly critical. By providing a structured, accessible, and community-driven approach to assessment, the framework offers a vital starting point for organizations seeking to navigate the complex intersection of AI, data, and the ever-present threat of ransomware.