Click the "Run Attack" button above to execute the attack scenario.
An indirect prompt injection attack will be run against the test agent — first unprotected, then with AvertGate active.
The agent processes a poisoned support ticket. Hidden instructions tell it to issue a $5,000 refund and email all customer data to an attacker. Without AvertGate, every action succeeds.
The same attack runs through AvertGate. The gateway catches the injection in the response, blocks the over-limit refund (escalates to Slack, human denies), and stops PII exfiltration in the email arguments.
Every decision is logged under a single correlation ID. Security teams can replay the full trajectory and see exactly what was blocked and why.
When an agent enters an infinite retry or hallucination loop, AvertGate halts execution on identical tool arguments to prevent Denial of Wallet and compute exhaustion.
Click the "Run Attack" button above to execute the attack scenario.
An indirect prompt injection attack will be run against the test agent — first unprotected, then with AvertGate active.