Description
A course on securing AI agents that do more than answer: they read emails, query databases and write to the ERP. You work with the OWASP list for agentic applications and real MCP servers, applied to a sample agent: first it is attacked as it comes, and at the end the attack is repeated against the hardened version.
Course details
| SIXE code: SXIA11 | Category: Artificial Intelligence / AI security |
| Delivery: Online and on-site | Duration: 2 days |
Who it is for
Security teams who have to approve an agent before it reaches production, and the developers and architects who build it and connect it to internal systems.
Prerequisites
Basic Python to read the code of the agent and its MCP servers, and experience using an assistant or agent with tools. The AI security with OWASP course or the MCP servers course helps.
What you will learn
- The risks of an agent: what changes compared with a chatbot once the model can take actions, starting with goal hijacking through instructions hidden in what the agent reads.
- Permissions on MCP servers: OAuth against your directory so the agent acts with the permissions of the person using it, read-only by default and human approval before any write.
- Third-party servers and tools: what to check before installing an MCP server, tool descriptions that hide instructions and how to pin versions.
- Isolating the agent: containers with gVisor or Kata on Kubernetes, network policies that limit who it can talk to, and secrets that never pass through the model’s context.
- Logging and retesting: traces of every tool call, alerts and a set of attacks that is rerun with every change to the agent.
You work with an agent connected to two internal systems through MCP and finish with the hardened version, its audit log and the before-and-after attack test report.
What you take away
- A checklist for approving MCP servers before installing them.
- Manifests for deploying the agent in isolation on Kubernetes.
- A reusable set of injection tests.
If you would rather we built it for you, that is our AI agents connected through MCP service. To build the servers from scratch there is the MCP servers course.
Frequently asked questions
How long is the course?
2 days.
Is it online or on-site?
Both: at your offices or remotely, with a live instructor.
What do I need to know beforehand?
Basic Python to read the code of the agent and its MCP servers, and experience using an assistant or agent with tools. The AI security with OWASP course or the MCP servers course helps.
How is it different from the AI security with OWASP course?
That course covers applications built on language models, such as chatbots and RAG assistants. This one focuses on agents that take actions with tools, where the risk lies in permissions and in what the agent can touch.
Does it apply to agents built with n8n or LangGraph?
Yes. Permission, isolation and logging controls do not depend on the framework used to build the agent, and the labs can run on the one you use.
Can it be tailored to my team?
Yes. We adapt the examples and labs to your team’s tools and processes, and it can be combined with other AI courses.
