Insidia, in about three minutes.

The long cut. It does not play until you start it.

Chapters

This support bot answered the question.

It also printed part of its own system prompt.

The instruction was hidden in the PDF. Nothing crashed.

An AI app is still an app, with a new kind of input.

Anything the model reads can carry an instruction.

AI checks and app checks usually run in different tools, on different days.

The risky path can start on one side and end on the other.

Insidia tests both layers in one run, on your machine.

It runs engines like garak, promptfoo, Nuclei, and ZAP, plus its own checks.

Every result names the engine that produced it.

Insidia only touches hosts you list.

Anything beyond localhost needs authorized: true, set by you.

A planted canary proves a leak. No guessing.

L1 is the baseline. Fast enough for every pull request.

Each row: the check, the result, the engine, the time.

A check that couldn't run is marked incomplete, never passed.

One HTML file. Opens offline.

The attack, the reply, and the proof.

It also lists what wasn't tested.

Treat uploads as data, not instructions.

Rerun checks again with the same config. Now it passes.

In CI, a failed policy blocks the merge.

Or hand it to your coding agent. You confirm the hosts.

Your AI agent is your newest attack surface.