The AI Shared Responsibility Model: a new challenge for security

For fifteen years, the shared responsibility model has played a crucial role in the context of cloud computing, allowing to answer the question "Who is responsible for this control?". This approach has avoided two main criticalities: the control that no one monitors because everyone assumes that the other does it and the control monitored twice in an irrational way.

However, the advent of generative artificial intelligence (Generative AI) has introduced new complexities. When an organization integrates a Large Language Model (LLM) into its processes, it entrusts a probabilistic component with the ability to interpret natural language inputs and produce content, code, or actions. The properties that make these systems useful are the same ones that complicate their security.

The new challenges of Generative AI

The behavior of LLM is adaptive, the output varies, the data sources change, increasing autonomy and the attack surface in ways that are difficult to predict. The consequences for the responsibility model are clear: it is not enough to adapt the table of the IaaS/PaaS/SaaS scheme, but it is necessary to recognize that AI introduces risk categories that did not exist in traditional software and that must be explicitly assigned.

Microsoft's three-layer model

The most well-known and operationally useful transposition of this model is the one formalized by Microsoft, which decomposes an AI application into three functional layers, each of which groups activities whose security normally falls to those who perform them.

  • AI Platform: The layer that provides applications with AI capabilities. Here reside the model, the training data, and the configurations that modify its behavior. The security system at this level is based on the presence of a safety system that filters potentially harmful inputs directed to the model and dangerous content that the model might produce as output.
  • AI Application: The layer that provides the service or interface that the user uses to access the model's capabilities. Security is entrusted to an application safety system that deeply inspects the content sent in the metaprompt and the interactions with plugins, data connectors, and other applications.
  • AI Usage: The layer that describes how AI capabilities are actually consumed. Security here is similar to that of any computer system, with a greater emphasis on user training.

Risk management

The first step to building the shared responsibility model in AI is mapping the three layers mentioned. This allows defining the new responsibilities, often unexplored areas in the context of traditional software. Organizations, in this phase, often make mistakes in assigning responsibilities.

AI does not simply shift responsibilities to the provider, but multiplies them and some remain with the organization, regardless of the model adopted. It is advisable to minimize the percentage of responsibility until you are able to support it.

Impact on the market and operational challenges

The adoption of the AI Shared Responsibility Model is radically transforming the landscape of corporate cybersecurity. According to a recent Gartner report, by 2026, 40% of organizations implementing AI solutions will suffer security incidents related to poorly defined responsibilities. This phenomenon is particularly evident in highly regulated sectors such as healthcare and finance, where regulations require precise traceability of automated decisions.

Companies are facing numerous operational challenges in the transition to this new paradigm. A survey conducted by ENISA revealed that 65% of the organizations interviewed encountered difficulties in mapping the three layers proposed by Microsoft, mainly due to the lack of internal expertise specialized in AI security. Many companies are responding by investing in targeted training for IT teams and collaborating with solution providers that offer responsibility assessment services.

The role of regulatory frameworks

The complexity of the AI Shared Responsibility Model requires a structured approach based on regulatory frameworks. The NIST AI Risk Management Framework (AI RMF) provides comprehensive guidance for identifying, assessing, and mitigating risks associated with the implementation of AI systems. This framework integrates the principles of the shared responsibility model, offering a methodology to align security practices with regulatory requirements.

In Europe, the entry into force of the AI Act is accelerating the adoption of these standards. Article 28 of the regulation requires providers of high-risk systems to clearly document the division of responsibilities between providers and end users. Companies operating in the Old Continent must therefore adapt their security policies to ensure compliance with these new provisions.

Strategies for risk management

To effectively implement the AI Shared Responsibility Model, organizations should adopt a gradual approach. A crucial first step is conducting a detailed analysis of their AI ecosystem, identifying all components that require protection and the related attack surfaces. This process requires collaboration between security teams, software developers, and corporate stakeholders.

An effective strategy involves adopting integrated security solutions that cover all three layers of the model. For example, in the AI Platform layer, it is fundamental to implement advanced filtering systems to prevent attacks based on prompt injection. In the AI Application layer, organizations should adopt real-time monitoring tools to detect anomalous behaviors in agentic systems. Finally, in the AI Usage layer, continuous training of staff is essential to prevent human errors that could compromise security.

Future trends and opportunities

The evolution of the AI Shared Responsibility Model is opening new opportunities for the cybersecurity market. Companies are developing innovative solutions to automate the management of responsibilities, such as resources available on the sites of Cybersecurity360 and specific frameworks like the OWASP Top 10 for LLM.

Additional resources

Editorial Note and Disclaimer

The guides and content published on GoYou are the result of independent research and analysis activities, for informational, educational, and in-depth purposes.

GoYou does not constitute a journalistic publication or an editorial product pursuant to Law No. 62/2001 and does not provide real-time information.

The GoYou project does not provide professional, technical, legal, or financial advice and disclaims all responsibility for the improper use of the information published.

In the Crypto sector, every investment involves risks: the reader is invited to always inform themselves autonomously before making any decision.