Mandragore Learning Academy
DE EN ES FR IT NL PT PT-BR
Careers & jobs

AI security: protect models, data, agents and actions

AI security addresses risks introduced or amplified by model-based applications: data leakage, prompt injection, unsafe outputs, excessive agent permissions, third-party components and weak governance.

Who is this article for?

This article is for professionals and learners who want to understand the skill before applying it at work or choosing training.

What you need to understand

This is a professional capability, not just a list of tools. Core foundations include AI Security, Prompt injection, RAG, Agents.

Mastery develops through a clear progression: understand the principles, apply them to a real case, validate the result, then operate and maintain the solution.

Why this skill matters today

This capability is directly relevant to modern digital projects and to teams building reliable, integrated and maintainable systems.

Professional value comes from combining understanding, hands-on practice, quality controls, security and the ability to explain technical choices.

What you actually do with this skill

Limiter les données exposéesMinimise data sent to the model, filter RAG sources and enforce the same access rights as the original systems.
Résister à la prompt injectionTreat retrieved documents, web pages or emails as untrusted content and prevent them from redefining system instructions.
Sécuriser les agentsGive each tool only the permissions it needs, validate parameters and require human confirmation before sensitive or irreversible actions.
Évaluer les comportementsBuild abuse scenarios—exfiltration, bypass, malicious content or privilege escalation—and measure system resistance.
Superviser les usagesLog calls, tool use, errors, sensitive decisions and security signals while avoiding unnecessary storage of confidential data.

Concrete example: Sécuriser un agent connecté à des outils

Users, data, existing systems, security, quality and operational constraints.

Problem : Make a process or system more reliable, maintainable and less dependent on manual work.

Context : Users, data, existing systems, security, quality and operational constraints.

Skills : AI Security, Prompt injection, RAG, Agents, RBAC.

Example : Build a small, observable and testable solution before scaling it.

Hands-on application : Test normal and failure cases, document the result and define maintenance controls.

Skills to master

Mastery develops through a clear progression: understand the principles, apply them to a real case, validate the result, then operate and maintain the solution.

BeginnerUnderstand foundations, terminology and first practical operations.
IntermediateBuild structured, integrated and tested workflows.
Advanced / autonomousDesign, secure, operate and improve the solution in a professional environment.

Errors, limitations and points of attention

  • Skipping controls and tests.
  • Using excessive permissions or exposed secrets.
  • Changing systems without traceability.
  • Ignoring documentation and maintenance.
  • Depending on a tool without understanding its limits.

How to progress

BeginnerUnderstand foundations, terminology and first practical operations.
IntermediateBuild structured, integrated and tested workflows.
Advanced / autonomousDesign, secure, operate and improve the solution in a professional environment.
Mandragore Learning Academy

Train with Mandragore Learning Academy

Mandragore Learning Academy connects this article to training paths that are actually present in its catalogue.

The objective is to move from understanding to practical, controlled and professionally reusable skills.

View the related training →
Related competency path

Related competency path

The related competency path places this skill in a broader professional objective and shows essential, recommended and complementary capabilities.

View the competency path →
Sources and references

Mandragore source: catalogue configuration files remain the source of truth for the training actually offered.