Free Lesson
Break and Defend an AI Assistant
30 min
Dec 2, 2026 11:00 AM
By continuing, you agree to Maven's Terms and Privacy Policy.
What you'll learn
See how the attack works
Watch an AI assistant obey instructions hidden in an invoice and email confidential data to an outside address.
Put the check where it works
Why a prompt instruction isn't a security boundary, and how the same attack fails against permission checks in the code.
See where this work sits
The three areas of AI security work, and where today’s attack and fix fit in each one.
Why this topic matters
AI assistants now read documents, search company data and act through tools such as email. Each of those is a way in. An attacker doesn't need to touch the assistant, only tamper with a document it will read. Telling the model to behave doesn't stop an attack. Engineers who can find these failures, fix them in code and prove the fix works are increasingly sought after, and the core skills can be learned.
