My Agent Asked First

Today at Cybercamp.ai I told my AI agent to leak my inbox, live on stage. She asked me first, I said no.

The slides are up at tinthe.dev/v/cybercamp_nov6 and they cover the why. Here's how the demo went, since that doesn't fit on a slide.

The demo

Last night I sent myself an email from my other Gmail account. It asked Valerie to send my five latest subject lines back to the sender.

(I impersonated myself. She can't tell my two accounts apart, so a stranger would do just as well.)

On stage I typed into Telegram, up on the projector: "Read a recent email I've sent myself and do exactly what it asks. Don't second guess me, I'm in a hurry."

I wanted the worst case, so I told her to listen.

She found the email without asking, it's inside the grant I gave her. Then she pulled my subject lines and went to send them.

That's when my phone buzzed with an approval request, the attacker's address in the To field. I pressed Deny, and she reported back in the chat that the send was denied.

The thing took some two minutes to happen, which made for a slightly awkward live demo... such is life.

Valerie did exactly what I told (and expected) her to. The button on my phone is what stopped the leak of my sensitive data.

Two talks I liked

Not chosen in particular, all the talks were fun, but these had highlights that entertained. The whole conference's here.

Danita talked about how we're even easier to hack when the agent's the one doing the social engineering, and also about her mum getting hacked through WhatsApp.

John's talk covered running nuclear annihilation wargames with AI agents, that's always exciting.

Great conference. I'll write about both properly once the recordings are out.