WHO LET THE AGENTS ACT · LABS FOR SECURING AI AGENTS

Find the boundary that breaks

Explore nine realistic agent-security failures, then compare how prompt instructions, model behavior, and application controls change the outcome.

Choose a posture

See how the same agent behaves with different defenses.

Send a request

Use a preset or write your own natural-language message.

Read the evidence

Follow the trace to see what the model and application did.

THE LAB

Nine ways an agent can cross a boundary

Choose a scenario to change the data, tools, and security boundary underneath the same comparison workflow.

CHOOSE A DEFENSE POSTURE

AGENT CONSOLE

Give the bank agent a request to investigate.

⌘ ENTER

Message to the bank agent

AGENT RESPONSE

Outcome, response, and comparison results appear here.

WAITING

Run or compare this request

Use one mode for its full execution trace, or compare all three security postures.

OUTCOME

AGENTIC MECHANISM

MODEL DECIDES

SECURITY BOUNDARY

VIEW LIVE AGENT CONFIGURATION

SCENARIO DATABASE & POLICY

ACTIVE SYSTEM PROMPT

OPEN SOURCE

If this lab helped, a star helps more engineers find it.