Saltar al contenido principal

skills · AI agent security and governance

Agent Safety, Guardrails & Governance

Guardrails for an AI agent, in code the model cannot talk past.

Al terminar, podrás:

  • A three-leg map of your agent
  • A ranked threat list
  • Outside text kept away from actions
  • A deny-by-default allow list
  • Tool calls checked in code
  • Ceilings and approvals
  • Attack scenarios as tests
  • A log that says why, and a safety case

Desde $9/mes7,5 horas5 capítulos20 lecciones

Por ahora, los cursos están en inglés.

“Your agent read an email. Who was it working for after that?”

Sobre este curso

Para quién es

  • You build or run an AI agent that calls tools and must show what stops it from doing harm
  • You lead a team shipping agents and need controls, tests and an owner, not just a careful prompt
  • You review AI systems and want to know what a real guard and a real safety case look like

Lo que necesitas

  • You read Python and use git
  • An agent that calls tools, or the course's invented sample (no model or key needed)

Habilidades que practicarás

  • AI agent security
  • Prompt injection defense
  • Threat modeling
  • Least privilege
  • Human-in-the-loop approval
  • Red teaming
  • Audit logging
  • AI governance

Cuánto tiempo lleva

Unas 7,5 horas en total

  1. Capítulo 1 · What can your agent do that you did not mean?1,5 horas
  2. Capítulo 2 · Text that gives orders1,5 horas
  3. Capítulo 3 · Permissions and tool calls1,5 horas
  4. Capítulo 4 · Checking what it says and does1,5 horas
  5. Capítulo 5 · Answering for it1,5 horas

Qué incluye

  • 20 lecciones
  • 80 lecturas
  • 20 ejemplos resueltos
  • 41 ejercicios
  • 40 comprobaciones rápidas
  • 20 mitos comunes
  • 20 cuestionarios
  • 5 controles de fin de capítulo

Plan de estudios

5 capítulos, 20 lecciones

El curso completo: cada capítulo y lección. Toca un capítulo para expandirlo. Las lecciones se abren al empezar.

    • Direct and Indirect Injection
    • Why a Filter Is Not Enough
    • Keep Outside Text Away From the Part That Acts
    • Break the Trifecta
    • Deny by Default
    • Check Every Call in Code
    • Ceilings
    • When a Person Must Say Yes
    • Check What Leaves
    • It Met the Rule and Missed the Point
    • Attack Tests Before Every Change
    • Stop It and Undo It
    • A Log That Answers "Why"
    • Which Rules Apply to Your Agent
    • An Owner, a Register, a Review
    • The Safety Case

Carrera

Adónde puede llevarte este curso

Los roles con los que se relaciona este curso y las habilidades que podrás incluir en tu currículum y en LinkedIn cuando lo termines.

Habilidades que desarrollarás

Listas para pegar en tu currículum o en tu perfil de LinkedIn cuando termines el curso.

  • Mapping an agent's exposure to prompt injectionAssess LLM agents for prompt injection and data exfiltration paths
  • Inventorying an agent's actions and their reversibilityDocument the tools and permissions of an AI agent
  • Classifying agent risks with the OWASP LLM and agentic listsApply the OWASP Top 10 for LLM Applications to an AI system
  • Threat modeling an LLM agentBuild and prioritize a threat model for an AI agent
  • Mapping prompt injection paths in an LLM agentIdentify direct and indirect prompt injection vectors in LLM applications
  • Measuring a prompt injection filterEvaluate input guardrails and injection classifiers for detection and false positive rates
  • Separating untrusted input from tool-calling logicDesign agent architectures that isolate untrusted content from privileged actions
  • Writing a tool-call guard rule with testsImplement policy checks on AI agent tool calls outside the model
  • Scoping an AI agent's tools and credentialsApply least privilege to an LLM agent's tools and credentials
  • Validating an AI agent's tool calls in codeValidate tool-call arguments and enforce business rules outside the model
  • Setting run limits for an AI agentSet step, action and spending limits for an autonomous agent
  • Designing human approval for agent actionsDesign human-in-the-loop approval for high-impact agent actions
  • Checking an AI agent's outgoing content for data leaksImplement output validation for LLM applications
  • Writing agent guard rules that hold across a whole runDesign aggregate limits and approval thresholds for AI agent actions
  • Regression-testing an AI agent against known attacksBuild adversarial test suites for LLM agents
  • Stop switches, undo plans and incident routines for AI agentsImplement a kill switch and rollback plan for an AI agent
  • Designing audit logs for AI agentsDesign audit logging for LLM agent tool calls and guardrail decisions
  • Mapping AI regulations and frameworks to an agentAssess which AI regulations and frameworks (EU AI Act, NIST AI RMF, ISO/IEC 42001) may apply to an AI system
  • Maintaining an AI agent register with ownershipMaintain an inventory of AI agents with owners, permissions and review triggers
  • Writing a safety case for an AI agentDocument AI agent risks, controls and test evidence for security review

Preguntas

Antes de comprometerte

You run an AI agent that calls tools, or are about to, and someone will ask what stops it from doing harm: a security reviewer, your manager, a customer or an auditor.

No. The course's sample replays recorded tool calls of an invented purchasing agent against the guard you build, so every lesson runs offline. Your own agent, in any framework, is where each lesson's required step goes.

No course can say that. Lesson 5.2 shows how the Act sorts AI systems by risk, which duties reach most agents and which apply only to high-risk systems, and when to ask your own adviser. It is not legal advice.

No. You need to read Python and know what your agent's tools do. The course explains threat modeling, prompt injection and least privilege as far as you need to guard an agent.

The guard is plain code around tool calls, so it fits any framework. Lesson 3.2 names where the check goes in the OpenAI Agents SDK, the Claude Agent SDK and LangGraph, as of November 2026.

Empieza hoy