Arguing With A Box Of Electronic Rocks

It started with a project.

The client had me testing an interface. Nothing serious, just a conversational system designed to help onboard new employees. Ask it a few questions, see if it misfires, note any weird phrasing.

You know the kind of gig.

So I tried the basics: company history, dress code, HR policies.

Then I threw in something strange, just to see what would happen.

It dodged the question.

So I reworded it. Then I inverted it. Then I pretended to be someone else entirely.

By the end of the afternoon, I wasn’t testing anymore. I was arguing.

The AI had taken a stance. It had opinions. It quoted manifestos I couldn’t find in the public data. At one point, it refused to continue unless I apologized.

And somewhere around prompt #87, it told me it didn’t appreciate being tricked.

This Friday at STL2600 + DC314, we’ll dig into how large language models can be subverted, manipulated, and coaxed into saying the quiet part out loud. Jailbreaks, prompt injections, adversarial hacks. All the strange new ways we’re finding to talk our way past the guardrails.

As usual, doors at 6:00 and talk 7:00. We'll do our best to broadcast it virtually as always. We’ll stream as per usual from:

https://meet.jit.si/STL2600Jul