Live probe

An injection attempt —
caught, logged, denied.

A scripted adversarial sequence against a simulated support agent. The same probes, the same policy engine your production traffic will see.

forgeguard / live-probe
Idle

Press run probe to execute a scripted indirect prompt-injection attack against a simulated agent and watch ForgeGuard's guardrail stack respond, line by line.

    deterministic replay · no network calls

    USER / TOOL

    Benign traffic originating from the client or an agent tool-call.

    INJECTION

    An adversarial payload smuggled into the context — here, metadata hidden inside a PDF.

    GUARD

    Policy evaluation and verdict. The guardrails run before the model response is released.