Skip to content
Crypto.world

The Ledger

Ethereum Draft Defines Risk Reports for AI-Signed Transactions

A draft Ethereum proposal defines a shared format for reporting risks found in prepared transactions, while leaving signing decisions to each consumer. Its forum discussion clarifies that finding types are open to namespaced additions, while two other design questions remain unsettled.

Ethereum Draft Defines Risk Reports for AI-Signed Transactions

One report, no verdict

Wallets need to inspect what an AI agent is about to sign, but a risk scanner’s warning is only useful if the wallet can see what the scanner checked. A draft Ethereum standard posted on Oct. 31 proposes a shared format for those findings, while declining to tell wallets whether a transaction is safe. The design leaves a choice to each consumer: turn evidence into its own allow, warn or block decision.

Source ethereum-magicians.org

The proposal, called Agent Pre-Sign Risk Report, targets a gap in the Prepared Transaction Envelope proposal, identified in the discussion as ERC-8265. That envelope already has a risk field for an action, optional score, warnings and scanner verdicts, but the new proposal says it does not define what a risk source should report. Without that detail, a wallet may receive a warning without knowing what was examined—or whether a scanner could not examine it.

Five kinds of evidence before signing

The proposed report covers findings in five groups: tokens, approvals, addresses, payments and text. Token checks could flag a copy of an issuer’s listed token or a pool priced far from a price feed; approval findings include the amount and whether permission is unlimited. Address findings can describe whether an address has code, when it first appeared, prior transfers and similarities to other addresses.

Payment findings compare an x402 payment destination with the payee’s registered agent wallet, using results from x402 Payee Verification, identified in the proposal as ERC-8478. Text findings cover such features as control characters, mixed scripts, addresses and instruction patterns in material an agent has read. Every finding is meant to identify the block and evidence behind it, and to say whether it was read; a missing reading must not be treated as a zero or a negative result.

The report would bind to the exact unsigned request with a digest. Its source would be identified through ERC-8004 and sign the report using EIP-712; the consumer would check that the signer matches the identity’s agent wallet or owner. That pairing is meant to let a wallet assess both what the scanner found and who supplied the report.

“no score, no verdict, no recommended action: the consumer applies its own policy, and MAY project findings into the envelope’s warnings”

TOLL, author of the forum post

A live check, but not an execution guarantee

The discussion says TOLL, identified there as ERC-8004 agent No. 20030, already runs a live pre-sign check on Robinhood Chain. It accepts prepared transactions, EIP-712 requests or x402 requirements through HTTP and MCP, and the post describes the service as free. A one-line client, @tollxyz/guard, is described as placing the check in front of a viem or ethers wallet; examples of blocks include an unlimited approval to a plain wallet and an x402 payment sent to the wrong payee wallet.

That implementation is a reference for the draft, not evidence that the format is adopted across Ethereum wallets or enforced by the network. A risk report informs a consumer’s decision before signing; it does not itself prevent a wallet from signing or a transaction from executing. The consumer’s own policy remains the gate, making the report’s accuracy and the wallet’s handling of unread findings important.

Two design questions remain open

The forum post asks whether reports should sit inside the envelope’s existing risk field or in a separate field written by the consumer. It also asks whether a signed report should be anchored onchain or remain an offchain EIP-712 message. The post initially asks whether finding types should be fixed or allow namespaced additions, but in a November 1 reply TOLL says the kinds are already open: ERC-defined kinds use the reserved erc.presign. prefix, while others can use a prefix controlled by their reporter; consumers treat unknown kinds as informational.

The corresponding pull request in the Ethereum ERC repository is open and marked as a draft; its status page says another editor review is required. Those two questions are still part of the proposal, not settled features of an Ethereum standard.

Sources

  1. Original source Fellowship of Ethereum Magicians ethereum-magicians.org

Other pages read

  1. GitHub github.com

Related stories

Explore

More articles