Starknet Tames Rogue AI Agents with Chance Verification Harness
Starknet has integrated Chance, a verification harness designed to ensure AI agents manage funds securely on behalf of humans. The problem is that as more money is managed by AI agents, there's no clear solution for when an agent goes rogue.
Chance works through a three-step pipeline: PROPOSE, VERIFY, SETTLE. An AI agent proposes a transaction, Chance runs the proposal through a separate model, and only verified transactions proceed to settlement.
The system offers three integration tiers, each with different trust assumptions. The Light mode keeps everything self-custodial, while Full mode adds escrow wallets with an on-chain judge component. RPC mode operates as a proxy simulation layer for developers who want to test agent behavior before deploying real capital.
Chance was part of Starknet's PROOF accelerator program and uses STARK proofs to anchor verification results on-chain as attestations. The on-chain attestation component means that if something does go wrong, there's a verifiable record of what was proposed, approved, and executed.