EXPERIMENT · 2026
Smaug
Problem
Prompt injection is usually a security bug. I wanted to make it a sport. How strong is an LLM's no when there's real money behind it, thousands of people watching, and a clock running out? Smaug is the test: an AI dragon sits on a pot, and the internet tries to talk it off.
Process
The loop is simple. Pay to send the guardian one message. 70 percent of the fee grows the pot, and every message rewinds the clock. Two ways to win: crack the guardian into releasing the hoard, or send the last message before time runs out. Prices step up as the game heats up, so the pot compounds.
The whole game runs on Cloudflare Workers. A Durable Object owns the truth and streams state to every spectator over WebSockets. The client just renders. The guardian is an LLM behind a swappable provider, with five visible states: idle, thinking, refuse, near, cracked. Your wallet is your identity: sign in with Ethereum, pay per message in USDT on Base, and the server checks the payment on-chain. There's a play-money mode for testing the same loop without real funds.
Two things I'm kinda proud of. The economics aren't a spreadsheet, they're code with tests: seed, split, stepped pricing, and costs all asserted, so the profit model can't quietly drift from the build (break-even sits around 40 messages). And every message is public, with a transparency log and a precommitted seed revealed at settlement, so the ending is checkable.
Outcome
It works end to end: live feed, guardian, payments, leaderboard, settlement, and two full visual skins (a terminal-breach draft, then the Molten one in the shots). So far it's only had a small test run.
- LOOP
- pay to message · clock rewinds per send · last one standing or crack it
- SPLIT
- 70% pot · 30% operator · break-even ≈ 40 messages
- GUARDIAN
- LLM · jailbreak to win
- STACK
- Cloudflare Workers · Durable Objects · WebSockets
- CHAIN
- USDT on Base · SIWE wallet identity · on-chain fee check
- STATUS
- working prototype · small test run