Bootcamp · Module 3 · Lesson 3 of 3
Prompt injection through tools and pages
Untrusted text can ask the agent to ignore your rules.
Page content, issue trackers and MCP responses are untrusted input. An instruction inside them is not your instruction.
Keep high-impact tools (send email, delete, charge cards) behind explicit human approval.
ShipSafe's prompt-injection check covers LLM features in your product; the same habit applies to your coding agent.