TABLE OF CONTENTS

Attackers redirect an agent's objectives through hidden instructions in documents, emails, or web pages

Agents apply legitimate tools in unsafe ways — deleting data, chaining APIs, or exfiltrating information

Delegation chains and cached credentials let attackers escalate access far beyond what was intended

Poisoned MCP tools, malicious plugins, and compromised registries inject malicious behavior at runtime

Prompt injection and unsafe code generation escalate into full remote code execution or sandbox escape