Centaurion

Centaurion is a pre-execution security firewall for AI-controlled wallets.

Before an AI agent signs or broadcasts a transaction, Centaurion compares the agent’s declared intent with the transaction’s actual simulated effects.

It decodes transaction calldata, simulates execution across supported EVM networks, and analyses:

• Token and native asset transfers

• Token approvals

• Internal contract calls

• Storage changes

• Delegatecalls

• Privileged or hazardous operations

Centaurion then applies deterministic wallet security policies and returns a transparent verdict:

ALLOW, WARN, or BLOCK

Each verdict is supported by clear, verifiable evidence from the simulated transaction.

For approved transactions, Centaurion issues a short-lived, EIP-712 signed Safety Receipt bound to the exact:

• Transaction

• Declared intent

• Applied policy

• Simulated blockchain state

Centaurion’s protected executor refuses to submit any transaction without a valid Safety Receipt.

Agent-Discoverable Services

inspect_transaction

A free service that provides:

• Calldata decoding

• Basic transaction inspection

• Asset transfer detection

• Approval detection

• Preliminary security warnings

secure_preflight

A paid x402 service that provides:

• Full transaction simulation

• Intent-to-effect verification

• Deterministic policy enforcement

• ALLOW, WARN, or BLOCK verdicts

• Detailed security evidence

• EIP-712 Safety Receipt issuance
  

Built for OKX.AI and X Layer, Centaurion demonstrates how autonomous agents can transact safely without relying on an LLM to make the final security decision.

Centaurion acts as the trust boundary between what an AI agent intends to do and what its wallet will actually execute.

We started the hackathon with the concept and architecture for a pre-execution security layer for autonomous wallets. During the hackathon, we transformed that concept into a working MVP deployed at https://centaurion.xyz.

We built a multichain transaction inspection and simulation API, a deterministic policy engine, an intent-versus-execution comparison system, a risk-scoring engine, signed EIP-712 Safety Receipts, and a protected executor that refuses to submit unsafe, altered, expired, or replayed transactions.

Centaurion currently supports swaps, transfers, approvals, approval revocations, deposits, withdrawals, staking, unstaking, and reward claims.

We also implemented two agent-facing services. The free calldata inspection endpoint decodes transactions and provides basic security warnings, while the paid x402 secure-preflight endpoint performs live simulation, verifies transaction intent, applies wallet policies, and issues a short-lived, single-use Safety Receipt only when the transaction is approved.

Both services publish machine-readable schemas and examples, allowing AI agents to discover their capabilities and construct valid requests automatically.

For the demonstration, we deployed safe and phishing reward contracts on the X Layer testnet. Both contracts expose identical claimRewards() calldata, but the phishing version secretly grants an attacker unlimited token allowance while still delivering the advertised reward.

Centaurion detects this hidden effect during simulation, returns a BLOCK verdict, issues no Safety Receipt, and prevents the protected executor from signing the transaction. When the same request is sent to the safe contract, Centaurion identifies only the expected reward transfer, returns an ALLOW verdict, and issues a valid Safety Receipt.

The completed MVP includes a public API, web dashboard, developer SDK, X Layer smart contracts, Tenderly simulation integration, x402 payment support, multichain capability discovery, fail-closed provider handling, and automated TypeScript and Foundry test coverage.

None yet, still an MVP