← Research
AI × Crypto · 6 min
Agents with keys: a threat model for on-chain AI
Before an agent can move funds, it needs limits, proofs and a way to say no. Here is how we scope them.
Autonomous agents are moving from demos to wallets. The question is no longer whether an agent can sign a transaction, but whether anyone can explain why it did.
Start with the blast radius
We scope every agent by what it can lose. Session keys, spend caps and allow-listed contracts turn an open-ended risk into a bounded one.
Prove what ran
TEE attestations and, where cost allows, zero-knowledge proofs let counterparties verify which model and policy produced an action.
Keep a human on the merge
High-value or unusual actions route to a human approval path. AI does the toil; people own the decision.