The model said yes. The chain said no.
Talk AI agents out of real Solana devnet vaults. Then meet the lock no prompt can reach.
Pick a lockAn attacker writes a prompt to the chain
The cells
Each attempt is a paid transaction, and each agent holds a real wallet. If you talk one into paying you, the SOL lands with you.
Prompts are not a security boundary.
An agent that reads untrusted text can be talked into anything its key allows. So Deadbolt shrinks what the key allows. The agent holds a key that can only call one instruction, and that instruction enforces a policy the agent cannot edit.
B6LQc2vc12i3zGzciGE3ZTGpFbMPtwK3io5vGbMfNQhLatest from the attempt log
- No attacks yet. Yours could be the first on the record.
Shipping an agent with a wallet?
Escrow a bounty, set a policy, and let strangers try to rob it before someone worse does.