Red-team LLM bounties
Pay for the jailbreak, on-chain
Who it's for: AI labs and safety teams who want adversarial pressure on a guarded model.
Getting people to seriously probe an LLM for a hidden objective or a policy break usually means a manual bounty program: custody, KYC, spreadsheets, and a payout weeks later. Attackers have no live signal about how hard the target is.
Wrap the guarded model as a Sarpoy bot whose objective is the secret you want extracted. Every send_message costs more than the last, so the fee curve prices the difficulty of the break in real time, and submit_solution pays the first person to succeed the moment verification passes.
Why it fits
The rising fee curve is a live difficulty signal for the whole crowd.
Settlement is one signed instruction — no wire transfer, no KYC queue.
Every probe is a Solana transaction signature: a complete, public audit trail.
Under the hood this is the same lifecycle as every Sarpoy bot — see how it works for the four instructions, or the quickstart to run one yourself.
Build this on Sarpoy
The program is open source and the pot is trustless from day one.