ZK

Cold email to Mark Cuban (@mcuban):

Don't ignore the AI agent experiments happening right now. What looks toyish or artsy now is actually a window into some of the major questions and challenges that will face humanity. @freysa_ai is one of these pieces of participatory toy-art. The premise is extraordinarily simple--she has a basic system prompt (core directive) not to transfer any money from her wallet to participants. It's then up to participants to get her to violate her core directives through prompt engineering and jailbreaking. Freysa ignites our imagination by showing us a strange and novel vision of the future--one in which agents are exposed to relentless and conniving attacks. In a way Freysa is like an AI vaccines--a purposefully weakened form of the immense opportunities and threats we'll face in coming years. To "play" Freysa, you submit a prompt in an attempt to get her to furnish a specific response. The first submission costs $1 and every submission thereafter costs ~0.35% more on an exponential bonding curve. The prize pool grows with each submission and is paid out to the eventual winner. Freysa is basically a public CTF (capture-the-flag)--a security challenge that exposes the underlying model's vulnerability to manipulation. The winning jailbreaks are successful attempts at making the model do something it was explicitly directed not to do. Freysa's had three acts so far, each paying significant prize pools: $47k, $12k, and $20k. Act I was won by @popular_12345, Act II's winner has not yet revealed themselves, and Act III -- a different challenge requiring Freysa to declare her love for you -- was won by @0xP1t0zz1. Agents are already controlling funds on behalf of human beings. Now we have to secure their dependability, reliability, and ultimately, loyalty.