📖 Quick glossary

• Prompt Injection : an attack that hijacks the AI via malicious instructions (OWASP risk #1).
• Direct / indirect injection : typed by the user / hidden in content the agent reads.
• Guardrail : automatic filter of inputs/outputs (toxicity, sensitive data, attacks).
• Lethal trifecta : sensitive data + untrusted content + external communication = dangerous agent.
• Least privilege : give the agent only the strictly necessary permissions.
• Stopping rule / Iteration cap : a step limit to avoid infinite loops.
• HITL (Human-in-the-Loop) : mandatory human validation before an irreversible action.
• Observability / Traces : detailed recording of the agent’s actions, reasoning and costs.
• Defense-in-depth : stacking several layers of protection (none being perfect alone).
• Execution history : native log (Make/n8n) of each run of a workflow — a ready-to-use trace.
• Service account/token : an identifier dedicated to the agent (≠ your admin account), with limited rights.