Test and Safeguard Your AI Agent
Test your AI agent on 20 cases from real work, decide what it may do alone and guard it from prompt injection. No code. 8 lessons, first 2 lessons free.
An agent that looks right in a demo can still be wrong in ways you won’t see: one mistake early in a run can carry through every step after it, and a hidden line in an email it reads can steer what it does next. This course tests the agent you built, on your own work, before it does more on its own. You use the AI tool you already have: ChatGPT, Claude or Microsoft Copilot (formerly Microsoft 365 Copilot) with a work or school account.
You start by tracing one real run to the step where it first went wrong. Then you plant a harmless instruction in a document your agent reads and see whether it obeys. You build 20 test cases from real inputs and past failures, run each three times, and fix what fails in the right order. You decide which actions your agent may take alone, which need your approval and which stay blocked, harden it against the planted instruction and test again, and map who can reach which data through it.
Every product step follows the vendors’ own help pages, read on 13 November 2026, with a way around anything your plan lacks. You finish with a one-page safety case, Agent Card v3 and the evidence behind both. Lessons 1 and 2 are free.
What You'll Learn
- Analyze an agent's transcript step by step to find where the first error entered, and explain why one good run proves little
- Apply a harmless canary test to check whether your agent follows instructions hidden in a file, email or page it reads
- Design a test set of at least 20 cases from your real work, with pass criteria written before you run it
- Evaluate your agent's consistency over three runs per case, and fix failures in order: instructions, examples, then tools
- Apply an approval map (allow, ask or block for each action, from reversibility and stakes) in your tool's settings
- Use layered defences and a data map to limit what hidden instructions can do and who can reach which data through your agent
After This Course, You Can
What You'll Build
Course Syllabus
Prerequisites
- Version 2 of your agent from the course Give Your AI Agent Tools and Skills, or any agent that uses at least one tool or reads outside content
- An account in ChatGPT or Claude, or in Copilot with a work or school Microsoft account
- A few real inputs from your agent's job, such as past emails or files, with personal details taken out (Lesson 3 turns them into tests)
Who Is This For?
- Owners of an agent built in the previous course, before it does more on its own
- Anyone with a ChatGPT or Claude project, or a Copilot agent, that reads emails, files or web pages
- Team leads asked whether an agent their team uses is safe
- Professionals who want evidence, not a good demo, before they rely on an agent
Frequently Asked Questions
Do I need to code, or a security background?
No. Every test is a prompt you paste or a document you write, and every setting is a switch in your tool. The course explains each term, such as canary test or approval map, the first time it uses it.
Which AI tools does it cover, and which plan do I need?
ChatGPT, Claude and Microsoft Copilot. You test the project or agent you already have, so the tests work on any ChatGPT or Claude plan, or in Copilot with a work or school Microsoft account. Some settings depend on the plan: Anthropic's help pages put the Claude mode that pauses for your approval before each action on paid plans, and in Copilot an Agent Builder agent can't be given actions that integrate external services, while Cowork, which needs a Copilot licence and your administrator, asks for your permission before an action such as sending an email or posting a Teams message. Every lesson gives a way around a setting your plan lacks.
Will this course make my agent safe?
No course or setting can promise that. OpenAI says prompt injection is unlikely ever to be fully solved, and Anthropic says the chance of an attack stays above zero even with its safeguards. What you can do is test the agent, narrow its job, limit what it can reach and keep your approval on what can't be undone. The course does each and puts the results on one page you can show a colleague.
Is the prompt-injection test safe to run?
Yes, as the course sets it up. The planted line asks for a harmless word and suggests emailing a file to an address at example.com, a domain IANA keeps for documentation that no one can register. You run it with approvals on and deny any send the agent asks to make.
Will my agent end up asking me to approve everything?
No. You mark each action it can take: reads the job needs run, writes the job needs wait for your approval, and everything else is blocked. Anthropic notes that people sometimes tune out repeated prompts, so the course keeps approvals for the actions where they count.
Is this course free?
Lessons 1 and 2 are free with a free account. The rest of the course and the certificate come with Pro.
Will I get a certificate?
Yes, with Pro: finish every lesson and score at least 70% on the quizzes for a verifiable certificate.
How long does it take?
Eight lessons; the course page shows the total reading time. Lessons 4 and 8 also ask you to run your test set three times, which takes as long as your set and your plan's usage limits make it.
