Security essentials · Free study companion
AI Agent Security: Prompt Injection to Payment Controls
Explore prompt injection, agent permissions and payment boundaries. Turn each lesson into a question to ask about your own AI workflow.
16 Whop lessons · 3 study checkpoints · 8 matching quizzes
Watch the course on Whop ↗
Checkpoint 1
Untrusted input and tool boundaries
Retrieved text, remembered context and tool descriptions are evidence to evaluate, not trusted instructions. Identify where untrusted input could change a goal, redirect a tool or leak information. Keep privileged policy outside that input path and verify the result with an observable check.
Watch the matching lessons (5)
Check your understanding
Agent Goal Hijack →
5 questions · Worked answers · No sign-in needed
Tool Poisoning Attack →
5 questions · Worked answers · No sign-in needed
Memory Poisoning →
5 questions · Worked answers · No sign-in needed
Practical challenge
Draw one untrusted-input-to-tool path for a toy agent.
Check your work
- Identify which text must not become policy.
- Name an enforceable boundary and an observable check.
Checkpoint 2
Wallet authority and operational safety
A model’s intention is not a security boundary. Token approvals, tool credentials and payment authority need enforceable limits outside the model. Repeated confirmation prompts can train users to stop checking, so approval design should preserve meaningful review of the actual destination and authority.
Watch the matching lessons (5)
Check your understanding
Approval Bypass →
5 questions · Worked answers · No sign-in needed
Confirmation Fatigue →
5 questions · Worked answers · No sign-in needed
Defense: Wallet Permissions and Signatures →
5 questions · Worked answers · No sign-in needed
Practical challenge
Design an external payment-approval policy for a toy agent.
Check your work
- Separate destination and spend-limit enforcement from model text.
- Show a rejected request and an allowed one.
Checkpoint 3
API payments and failure handling
Payment workflows need clear authority and state transitions. Separate request, authorization, settlement and retry; an ambiguous response is not proof that nothing happened. Verify the counterparty and enforce budgets outside the model, with a response plan for untrusted outputs and unavailable dependencies.
Watch the matching lessons (4)
Check your understanding
Asynchronous Settlement →
5 questions · Worked answers · No sign-in needed
Incident Response →
5 questions · Worked answers · No sign-in needed
Practical challenge
Write a state machine for a paid API request with a timeout.
Check your work
- Distinguish attempted, authorized and settled actions.
- Explain how you investigate an ambiguous result before retrying.
Explain the decision, not just the answer.
After each checkpoint, write one assumption you changed and one test you would add. Bring a sanitized example or a specific question to the free Academy community.
Find your way around the community →