⚡️Jailbreaking AGI: Pliny the Liberator & John V on Red Teaming, BT6, and the Future of AI Security
Source
youtube.com
Author
Latent Space
Date
Context
A jailbreakA prompt crafted to make a model ignore its own guidelines — usually through roleplay, hypotheticals, or encoding rather than a direct request.Full definition → test can involve an entire conversation. The adversarial AI testers in this interview distinguish template-style attacks, where one prepared input is dropped into a model, from what they call soft jailbreaks. A soft jailbreak is a slow, multi-turn exchange that steers the model while avoiding the triggers that would shut the conversation down, similar to a crescendo attack. In their account, the attack lives in the sequence of turns rather than in any single prompt. An evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → that sends only isolated requests would therefore miss this style of attack, which gives evaluators a concrete reason to test whole conversations.
Terms in this piece · Glossary
jailbreak — A prompt crafted to make a model ignore its own guidelines — usually through roleplay, hypotheticals, or encoding rather than a direct request.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
red teaming — Deliberately attacking your own AI system to find what makes it fail before someone else does.
videoBreaking AI to Fix It: Ian Webster's Journey from Discord's Clyde to Promptfoo's $18M Series ALatent Space
videoWhen AI Agents Run Businesses — Lukas Petersson and Axel Backlund of Andon LabsLatent Space
video⚡️OpenCode: Claude Code but Open Source, with Any Model, and frontier TUI - with Dax Reed (@thdxr)Latent Space