Securing and governing an AI agent in production

Objectives

• translate an agent's written scope (accessible data, authorised tools, forbidden actions, escalation cases) into technical controls: permissions, allow-lists, access segregation
• test an agent against the attacks specific to it (direct and document-borne prompt injection, data extraction, tool hijacking) and fix the flaws found
• log every action of the agent in a usable way (who, what, when, on which data, with what result), with a defined retention period
• put human validation in place for irreversible actions, input and output guardrails, rate and cost limits, and an emergency stop
• write the response plan for an incident involving an agent (wrong answer circulated, leak, behaviour drift) and run it during an exercise

Audience

Developers and AI engineers who have taken AGT-301 or who run an agent in production, security engineers, DevOps, technical teams taking over an agent built by a provider.

Prerequisites

Have taken AGT-301 or maintain an agent in production, and program in Python. Checked by a positioning test. Technical requirements: one laptop per participant, 16 GB of memory recommended, with Python 3, Docker, Git and a code editor installed before the first day (procedure provided with the invitation); a deliberately vulnerable training agent, a cloud environment, API keys and tracing tools provided by Yolea, credits included; for remote delivery, a stable connection of at least 10 Mbit/s, microphone and camera; the participant's own agent is accepted for the workshops if it can be deployed in the training environment.

Programme

• 25% inputs, 75% workshops
• Alternating attack and defence on the training agent, each flaw exploited and then fixed
• Building the action log and the supervision dashboard
• Timed incident exercise on the second afternoon
• Practical cases inspired by company situations (a WhatsApp sales agent with access to the catalogue and orders, an agent that pre-assesses files)
• Tools of the current edition: LangGraph or Pydantic AI, MCP with permissions, LangSmith for traces, input and output guardrails
• Materials handed out: a reference repository, a reusable battery of attack tests, a technical scope template, a log template, an incident response plan, a glossary
• A personal action plan
• Follow-up point at three weeks

Format

In person recommended, hybrid possible; in-company on your own agent, or open enrolment (inter-company); 8 to 12 participants; 14 hours, 2 consecutive days of 7 h; remotely, 4 sessions of 3 h 30 over two weeks.

Assessment

Beforehand, a positioning test. During the course, a check quiz and a configuration review. At the end, a practical challenge (secure a given agent against an unknown battery of attacks, log and handle a simulated incident, scored on a grid), an immediate feedback survey and a named certificate of course completion. At three weeks, a follow-up point.

The last exercise is an action plan: three real tasks that you leave ready to equip with tools.