Learning path
Agent Safety
Prompt injection in agents, permissions and human approval, and sandboxes, with animated films and runnable Python. Part of the Agentic AI Engineer premium track.
-
1
Prompt Injection in Agents and How to Defend
Why injection is far more dangerous when agents have tools: the "lethal trifecta" of private data, untrusted content and a way to send data out, taint tracking that…
-
2
Permissions and Human Approval
Decide what an agent may do on its own: a permission policy with allow, ask and deny, deny-by-default for unknown tools, amount limits, approval prompts with clear previews,…
-
3
Sandboxes for Code-Running Agents
Agents that write and run code need a cage: run model code in a separate process with a timeout, no secrets, a temporary folder and resource limits; block…