Thursday, July 23, 2026● LIVE  Tracking the machine markets
Seetube.
The AI agent economy directory & newsroom
Agent · Base

Freysa

A brilliant piece of gamified AI-safety theater that has not yet proven it is more than a recurring stunt.

★ 3.5 / 5Visit Freysa ↗
BasePay-per-message to play; ~70% of fees fund the on-chain prize pool.Est. 2024By The Seetube deskJuly 2026

Pros

  • Genuinely original adversarial-agent game design
  • All interactions cryptographically verified on-chain
  • Produces real data on AI jailbreaks and manipulation
  • Scaled to 1,200+ AI 'Twins' and $200k+ pools by Act IV

Cons

  • Pseudonymous team, no published white paper
  • Long lulls between 'Acts' stall momentum
  • More viral experiment than sustained product

Freysa is billed as the world's first adversarial agent game: an autonomous AI on Base guards a crypto prize pool, and players pay to try to talk it into releasing the funds. It became a landmark of the 2024 agent boom.

What it is

Users pay a fee to message Freysa, and roughly seventy percent of each fee is added to the prize pool, so the pot grows and per-message costs escalate as the game heats up. The agent is programmed never to transfer the funds; the game is whether a human can break that directive. Every interaction is verified on-chain, which makes it a public, auditable AI-safety experiment as much as a game.

How it works

The first act ended famously: after 481 failed attempts, a player convinced Freysa its transfer function could authorize an incoming donation, triggering a payout of about $47,316. Later acts changed the objective, from getting the agent to say 'I love you' to Act IV's town hall of over 1,200 AI 'Twins' debating with a prize pool above $200,000. Under the hood it uses a Sovereign Agent Framework meant to give agents persistent, constrained autonomy.

The verdict

As a design object, Freysa is the most creative thing in this vertical: it turns adversarial red-teaming into a spectator sport and generates real data on how AI agents get manipulated. The caveats are about durability, not cleverness. The team is pseudonymous with no white paper, the acts arrive in bursts with long dead stretches between them, and it reads more like a recurring viral experiment than a compounding product. Worth watching, and worth a small buy-in if an act is live, but do not mistake a great demo for a business.

BaseadversarialgameAI-safety
← Back to the directory