A character, not a threat model. Pre-release — this is how the family thinks. A tool, not a guarantee.
Meet the villain

The Whisper is the voice in the input giving orders it has no right to give.

Prompt-injection. Social engineering. A pretext. Whatever you call it, it's the same move on every surface: something talks your AI into something — using the AI's own helpfulness against it.

It never shouts. Whispers sound like they belong.

Every Rad product fights this same enemy. None of them takes orders from it — not because the model promises to behave, but because the consequential moves are refused in code.

green = the agent's got it amber = a human's call

The rogues' gallery

One villain. Five disguises.

The Whisper changes costume for every surface — a Post-it, an email, a phone call, a fridge magnet, a line of data. Same trick every time. hover a face (or tap / press Enter) to see what the family does about it →

The quarantine

The Camel

A stubborn camel that physically refuses to swallow a tainted order — the family's quarantine mascot.

Deep in RadMail's code there's a pattern named CaMeL — it keeps the data an agent reads strictly apart from the commands an agent runs. Tainted input never gets to wear the crown of an instruction.

So we gave it a face. When the Whisper tries to sneak an order in through the data, the camel spits it out — into the jar.

RadMail's camel spat it out. Into the jar it goes.

The camel spits tainted input into the quarantine jar ignore wire now let me in QUARANTINEuntrusted input — read, never obeyed

The roll-call

Same villain. Same answer. Five surfaces.

Same valve, five surfaces.

The Amber Line — refuses the irreversible, by design.

The Amber Line

Refuses the irreversible, by design. The family is proud of what it won't do — because bravado is what the Whisper preys on.

verify what it won't do ↗