Complete guide to security in AI development environments

Informatec Digital » Resources » Complete guide to security in AI development environments

Organizations developing or integrating artificial intelligence solutions are facing a completely new regulatory landscape in 2025. The European Artificial Intelligence Regulation (RIA) and the supporting guidelines published in Spain have changed the rules of the game for those who design, train, deploy, and monitor AI systems, especially when they are considered high-risk.

Faced with a complex, technical, and still-evolving legal framework, many companies are urgently wondering where to begin. The 16 guides and checklists published by the Spanish Agency for the Supervision of Artificial Intelligence (AESIA) have become the practical reference point for implementing the RIA (Research and Innovation Analysis) in the daily reality of SMEs, startups, and large companies that want to innovate without getting bogged down in regulatory complications.

The Ministry for Digital Transformation and Public Administration, through the State Secretariat for Digitalization and Artificial Intelligence (SEDIA), has opted for a highly pragmatic approach. Before the European regulation became fully enforceable, an AI Sandbox was launched to test the regulation with twelve real-world , high-risk systems from sectors as sensitive as essential services, biometrics, employment, critical infrastructure, machinery, and in vitro diagnostic medical devices.

In this testing environment, the Ministry team worked closely with the selected companies, providing high-level training and public consulting. The objective was twofold: to help the organizations understand what the RIA will require in practice and, at the same time, to identify ambiguities and frictions in the standard when applied to algorithms and processes already functioning in the real world.

A wide range of companies participated, from firms specializing in biometrics and advanced analytics to providers of digital health and employment solutions. Key supervisory authorities such as AESIA, the Spanish Data Protection Agency (AEPD), the Bank of Spain, and the State Aviation Safety Agency (AESA) also took part in the project , along with a panel of regulatory and technology experts.

This collaborative work resulted in the conclusions now presented in these guidelines. While these documents are not legally binding and do not replace the harmonized technical standards that will be forthcoming, they do offer a practical, organized, and realistic interpretation of the RIA's obligations for high-risk AI systems.

Thanks to this experience, Spain has earned a leading international role. The AI ​​Sandbox is cited as a best practice in the Report for Europe's Digital Decade and is viewed by other countries as a model for anticipating regulatory compliance , reducing uncertainties and costly mistakes along the way.

The 16 AESIA guidelines: structure, approach and objectives

AESIA's publications are organized into a coherent set of 16 guides plus several checklists. Their main purpose is to help organizations identify, understand, and comply with the obligations of Regulation (EU) 2024/1689 , focusing on systems classified as high-risk.

These guidelines are specifically aimed at SMEs, startups, and large companies that develop or deploy AI in regulated sectors. They do not replace regulations or standards that will be adopted, but serve as a support tool for designing innovative, reliable AI systems that respect fundamental rights —essential for both fostering and unchecking innovation.

The collection is divided into several functional blocks. On the one hand, it includes two introductory guides of an informative nature, and on the other, thirteen technical guides focused on the specific requirements of the RIA (risk management, data governance, transparency, cybersecurity, quality, records, human oversight, etc.).

In addition, a specific manual for checklists has been published, and sample checklists are provided in Excel format. These files allow for self-assessments of the level of compliance of high-risk systems and the planning of an orderly adaptation roadmap , instead of improvising last-minute fixes once the regulations are fully in force.

An important feature is the modular structure of the documents. The guides are designed so that each organization can delve deeper in layers according to its size, sector, and level of AI maturity ; this prevents an SME from being overwhelmed by the same level of detail that a large multinational with dozens of production models would require.

Introductory guides: understanding RIA without going crazy

The two introductory guides provide the conceptual foundation to avoid getting lost in the regulatory jargon. The first is a general introduction to the RIA, describing its scope, risk categories, and the different roles involved (supplier, deployment manager, distributor, user, etc.).

It clarifies issues that often cause confusion for companies, such as what truly constitutes a high-risk AI system, which uses are directly prohibited, and what obligations each stakeholder in the value chain has. It also details how the RIA relates to other relevant European regulations, such as data protection, product security, and financial services , which is crucial for avoiding duplication or loopholes.

The second introductory guide takes an even more practical approach, combining regulatory explanation with a case study. It shows how to identify whether a particular system falls into the high-risk category and what basic steps it would need to take to comply with the regulations , from the design phase to monitoring once it is on the market.

Both documents include a detailed glossary, which is very useful for non-legal teams. Defining terms such as "human oversight," "reasonably foreseeable risk," and "systemic risk" precisely prevents many internal misunderstandings between technical, legal, business, and cybersecurity departments.

These context guides have become, in practice, the go-to resource for managers, DPOs , CISOs, innovation leaders, and IT teams who need a clear overview before delving into the more technical aspects. They function as a general map, directing the reader to the specialized guides they actually need and preventing them from wasting time reviewing documents that don't apply to their situation.

Technical guidelines: grounding RIA requirements in the AI ​​lifecycle

The largest section of the AESIA collection consists of thirteen technical guides, each dedicated to a key RIA requirement for high-risk systems. These guides get down to brass tacks and translate legal mandates into verifiable processes, controls, and evidence that organizations can implement in their AI projects.

The conformity assessment guide explains how high-risk AI systems should be prepared to pass the assessment procedures set out in the regulation. It describes the required documentation, how internal and external audits are organized , and the role of notified bodies in certifying that a model meets the essential requirements.

Another guide focuses on the quality management system. Its purpose is to ensure that the metrics and objectives defined during system design remain stable throughout development, validation, and operation , and that any deviations are properly recorded. It emphasizes the importance of maintaining evidence of all validations and tests performed, not only for compliance but also to be able to defend the system's robustness to third parties.

The risk management guide translates the RIA's risk-based approach into an operational methodology. It includes guidance for identifying, analyzing, assessing, and mitigating risks related to health, safety, and fundamental rights , both in the intended use of the system and in reasonably foreseeable uses. It also addresses the need for periodic reviews as data, context, or the technology itself changes. To plan practical measures, it is recommended to use self-assessment tools that provide guidance on mitigation priorities.

Human oversight is covered in a separate, comprehensive guide. This guide details the design, interface, and organizational processes necessary for those monitoring the system to detect anomalies, intervene promptly, and avoid blindly relying on automation . It proposes, for example, human override mechanisms, clear alerts, and well-defined escalation procedures when something seems amiss.

The guide on data and governance is especially relevant for complex models. It describes the requirements that datasets used to train, validate, and test AI systems must meet : relevance, representativeness, absence of unjustified bias, quality, integrity, and traceability. Furthermore, it addresses the need for governance policies that control who can modify the data, how sources are documented, and how errors detected afterward are handled.

Another technical document focuses on transparency. It proposes measures to enable both providers and users of AI systems to comply with their reporting obligations : from making technical documentation accessible to integrators to ensuring clear communication with people affected by automated or semi-automated decisions.

The accuracy guide provides guidance on how to select, measure, and monitor model performance metrics. It is not limited to statistical accuracy; it also addresses issues such as temporal stability, subgroup performance, and behavior under boundary conditions , in order to avoid misleading results that could negatively impact specific groups.

The system's robustness is also addressed in its own document. This guide outlines the measures that must be taken to ensure fault tolerance, resilience against adversarial inputs or changes in the environment , and the ability to continue operating within safe parameters even if some components fail or data degrades.

Cybersecurity is addressed in a dedicated guide that outlines the technical and organizational measures necessary to protect AI systems against malicious attacks or manipulation. It covers aspects such as data pipeline security, model exfiltration protection, access control, and incident response , aligning these requirements with the organization's broader information security vision. For more in-depth information on specific threats, see Cybersecurity Threats and Their Impact on AI.

The logging guide defines which logs and evidence should be retained in high-risk AI systems. It includes guidance on which events to record, how long to retain them, and how to ensure their integrity and availability so they can be used for audits, incident investigations, or claims from affected individuals. The use of appropriate encryption mechanisms is recommended for protecting the integrity of the logs.

Another guide focuses on post-market surveillance. This document describes the processes necessary to monitor system performance after deployment , gather user feedback, detect incidents, adjust the model if new risks emerge, and report to the authorities when appropriate.

The technical section is completed by the incident management guide, which explains how to report serious incidents related to high-risk AI systems, and the technical documentation guide. The latter outlines the content and structure requirements for the technical dossier that will demonstrate compliance before the system is put on the market or into service.

Checklists: From Paper to Action in Business

Alongside the guidelines, AESIA has published a manual of checklists and examples of structured checklists in Excel format. Its purpose is not theoretical, but rather to serve as an operational tool so that organizations can assess the level of compliance of their high-risk AI systems themselves.

These lists follow the same logic as the requirements guides. For each of the twelve major RIA requirements, a set of questions, evidence, and fields to complete is included , which help to conduct a structured and consistent self-assessment. The file includes information tabs with instructions for use and context, as well as worksheets for entering specific information about each system.

The idea is that companies shouldn't just "tick boxes," but rather use the tool to build a genuine adaptation plan. The self-assessment should result in a realistic roadmap that prioritizes risks and measures based on impact and effort —something especially critical for organizations with limited resources.

This approach is especially useful at a time when deadlines seem long but the work is intense. The most demanding obligations of the AI ​​regulation will begin to apply in August 2026, but reviewing internal processes, data models, and governance doesn't happen overnight , so those who get ahead will have more leeway and less improvisation.

Furthermore, the systematic use of checklists makes it easier for very different teams (legal, cybersecurity, IT, product, data) to speak the same language. When everyone works with the same criteria and levels of evidence, internal noise, delays, and unnecessary redesigns are reduced , which in AI projects usually translates directly into cost savings and fewer headaches.

Code of good practices for general-purpose AI models

While Spain was making progress with its Sandbox and the AESIA guidelines, another key instrument was being developed at the European level: the Code of Good Practice for General Purpose AI (GPAI). The European AI Office coordinated a participatory process with nearly a thousand stakeholders to develop this code, which complements the obligations of the AI ​​Act for providers of general purpose models.

General-purpose AI models are those capable of performing many different tasks and that underpin a multitude of systems and applications. Some of these models can generate systemic risks if they are very powerful or widespread in the market , which is why European legislators have defined specific requirements for transparency, copyright, and security.

The Code, submitted to the European Commission on July 10, 2025, is a voluntary tool, but with clear advantages. Suppliers who adhere to the Code can use it to demonstrate compliance with the obligations of Articles 53 and 55 of the RIA , thereby reducing the administrative burden and improving their legal certainty compared to ad hoc solutions.

The document is structured into three main sections: transparency, copyright, and security/protection. Regarding transparency, the Code provides a standard model documentation form, which collects key information on capabilities, limitations, training data, testing, and validation . This information must be kept up-to-date and available to authorities and downstream providers upon request.

Regarding copyright, the Code translates European intellectual property legislation into the context of AI. It recommends specific measures such as limiting web crawling to legally accessible content, respecting copyright reservations through robots.txt or metadata, excluding infringing domains, and establishing clear internal copyright policies , designating responsible parties and complaint channels.

The third section addresses general-purpose AI models with systemic risk, meaning those whose high-impact capabilities could have a significant effect on the EU market. It proposes principles of comprehensive lifecycle management, proportionality to the level of risk, innovation in security, and application of the precautionary principle , especially when there is still insufficient scientific data to fully assess certain hazards.

The Code also focuses on SMEs that develop advanced models. It advocates for simplified compliance pathways for small suppliers, provided the expected level of protection is maintained , and details obligations such as the notification of serious incidents or the documentation of preventive actions before those incidents occur.

This instrument is not static. The text itself provides for its review at least every two years, so that it can be updated based on technological developments, the emergence of new risks, or the experience gained from applying the RIA in the field.

Template for summarizing training data

In parallel with the Code of Good Practice, the European AI Office has developed a specific template for summarizing the training data that providers of general-purpose AI models must make public. This obligation, stipulated in Article 53.1(d) of the RIA, aims to strengthen transparency and respect for copyright in model training.

To develop this template, the Office gathered input through a multilateral consultation, receiving over 430 responses from various stakeholders. The preliminary ideas were also reviewed within the framework of the Code of Good Practice process, with input from suppliers, experts, representatives of Member States, and the European Parliament , before its formal adoption in July 2025.

The template is closely linked to the Code's transparency chapter and seeks a delicate balance. On the one hand, it requires a summary detailed enough for third parties to reasonably understand what types of data have been used ; on the other, it respects the confidentiality of sensitive business information and does not require the disclosure of every specific element of the dataset.

In practice, this tool fosters a more mature documentation culture among general-purpose model providers. Those training large models with enormous volumes of data need to clearly understand the data sources, the licenses governing their use, and how all of this will be reflected externally , both to manage legal risks and to build trust within the ecosystem built upon those models.

Again, the emphasis is not merely formalistic. Transparency in training data is not just about compliance; it also allows for honest discussion of potential biases, reasonable use limits, and realistic performance expectations when these models are integrated into solutions for critical industries or large-scale end users.

AI tool audit and strategic vision for executives

Beyond purely regulatory obligations, many organizations are considering how to evaluate the true return on investment of their AI initiatives in 2025. It's not enough to simply deploy intelligent models and agents; it's necessary to audit whether they generate real value, whether they are secure, and whether their total cost of ownership makes sense compared to other technological alternatives.

A well-designed AI tool audit allows CEOs, CIOs, and other executives to separate marketing hype from tangible results. This type of analysis reviews how well each use case aligns with business objectives, measures expected benefits in productivity, error reduction, or revenue generation, and analyzes technical and regulatory risks such as bias, cybersecurity vulnerabilities, or RIA non-compliance.

In this context, very specific metrics become relevant: cycle time saved, reduction in operating costs, increase in conversions, accuracy of the model in production, cost per transaction, and savings attributable to automation. Risk indicators are also assessed, such as vulnerabilities detected in security tests, level of regulatory compliance, and the likelihood of excessive dependence on a single vendor (vendor lock-in).

Companies that complement AI with business intelligence and visualization solutions, such as dashboards using Power BI tools, gain a much clearer view of its impact. Centralizing performance, cost, and risk data in dashboards that are easy for management to understand facilitates evidence-based decision-making , rather than relying on intuition or fleeting technological trends.

In parallel, integration with cloud infrastructures (AWS, Azure, or others) and accurate estimation of the total cost of ownership are key components of the analysis. A thorough understanding of the costs of training, inference, storage, and maintenance in the cloud, as well as the associated security mechanisms, is fundamental for AI to truly support the strategy and not become a costly expense that is difficult to justify. For decisions regarding cloud data backup and protection, consult comparisons of cloud and on-premises backups.

In short, AI auditing is no longer a luxury but a strategic necessity. Those who combine a rigorous approach to RIA compliance with an honest measurement of technology ROI will be better positioned to leverage artificial intelligence sustainably , both in terms of economics and reputation and social trust.

This entire ecosystem of guidelines, voluntary codes, transparency templates, and auditing practices will shape a new framework for security in AI development environments by 2025. Organizations that begin now to internalize these tools, use them systematically, and integrate them into their technology governance will have a greater chance of turning regulation into a competitive advantage , demonstrating that it is possible to innovate in AI responsibly, securely, and in alignment with human rights.