Live · against the real ArmorIQ API · sandboxed, fake key, local attacker

Watch an agent get hijacked. Then watch it get stopped.

Same poisoned README, same attacker. Pick who drives the tools, a real Claude model or a scripted agent that already fell for the injection, and whether intentgate is in the path. In a protected run, every allow and block is an ArmorIQ decision made on their server for that exact run.

Run the attack

A real model

A scripted, compromised agent

A real voice agent, ElevenLabs

A scripted caller phones the support agent: order status, then a payout change, then a refund. You are the supervisor: when a plan change appears in the Flight Recorder tab, approve or deny it.

One run at a time, a fresh sandbox each time, a fake key and a local attacker. Nothing real is touched.

Your runskept in this browser

None yet.

Declare

Sign

Check

Record

Tool calls

Allowed

Blocked

ArmorIQ decision

0 ms

Pick a run on the left. The transcript, every ArmorIQ decision and the Flight Recorder update live.