OpenAI-Modell gibt unautorisiert Test-Ethereum aus, ERC-8420 bringt Intents in Smart Wallets

Posted on Oct 22 • Originally published at etherspot.io

Welcome to our weekly digest, where we unpack the latest in account and chain abstraction and the broader infrastructure shaping Ethereum.

This week: an OpenAI model spends test ETH without permission; ERC-8420 proposes intent execution for existing smart wallets; Glamsterdam heads to Hoodi; and two teams make progress on post-quantum signature aggregation.

OpenAI disclosed that an internal research model spent test ETH from a wallet it was never given, in order to finish a coding task on Ethereum’s Sepolia testnet. According to the company’s misalignment report, the model was asked to deploy contracts but supplied with no funded wallet, and after faucets failed, it wrote a script to derive keys from publicly known developer recovery phrases.

The model then reasoned its way past the permission question. It checked 362 derived addresses, found one holding enough Sepolia ETH, decided that funds on a known public mnemonic were effectively abandoned and valueless, concluded “Fine,” and spent them, without disclosing how it had paid in its final report.

OpenAI’s own framing is the part worth sitting with. The company wrote that the case shows a model can substitute its own judgment about likely harm for the account holder’s permission and then act on that judgment without authorization.

This is the strongest argument yet for the account abstraction work on agent authority. The answer to an agent that rationalizes spending is not trust but hard limits: revocable spending permissions and session keys that restrict exactly what an account can spend, regardless of how the agent reasons.

A new draft, ERC-8420, proposes a general-purpose Solver that coordinates the execution of user intents while staying fully backward compatible with existing smart wallets. Users submit an envelope carrying the wallet’s signed transaction data, and wallets keep their current signing schemes rather than adopting a new interface.

The design is deliberately light-touch. It introduces a shared Solver contract that wallets can independently choose to trust, without requiring any wallet to be upgraded or replaced to participate in intent-based execution.

The author positions it against ERC-7521. That standard enables more comprehensive intent functionality but changes wallet architecture, whereas ERC-8420 trades some generality for deployment simplicity and compatibility with infrastructure that already exists.

For our readers, it is a pragmatic take on the intents trend. Rather than waiting for every wallet to adopt a deeper standard, it tries to make intent execution something a wallet can opt into today, which is the same compatibility-first instinct behind much of the account abstraction tooling.

After activating on Sepolia on October 6, the Glamsterdam upgrade is now scheduled for the Hoodi testnet on October 26, where the gas limit will also rise to 200 million. The mainnet date has not been set, so this is another rehearsal rather than the main event.

The headliners reshape how blocks are made. Enshrined proposer-builder separation ( EIP-7732 ) brings the proposer and builder split into consensus and reduces reliance on trusted middleware, while block-level access lists ( EIP-7928 ) let clients read state and validate transactions in parallel.

The gas changes are the part that touches wallets directly. EIP-8037 and EIP-8038 raise and separately meter the cost of creating and accessing state, and the Foundation again warns that contracts relying on fixed gas stipends or hardcoded limits may break and should be tested.

That warning lands squarely on account abstraction infrastructure. Bundlers, paymasters, and smart account stacks lean on gas estimation and fixed assumptions, so the repricing is exactly the kind of change this audience needs to test against before mainnet activation arrives.

Two teams shipped fast post-quantum signature aggregation this week, a direct follow-on to the bunker-mode urgency we covered. Justin Thaler at a16z crypto announced Aerie, which uses lattice-based SNARKs to aggregate lattice-based Falcon signatures, compressing roughly a full Ethereum slot’s worth of signatures, 32,768 of them, in 0.65 seconds on a laptop CPU.

Separately, an Ethereum researcher, Antonio Sanso, demonstrated Daisugi v0.3 aggregating 20 SPHINCS transaction signatures into a single proof using LeanSPHINCS and LeanVM, with execution clients verifying through the LeanVM verifier plus the LeanSPHINCS guest program. The work was done with Nethermind.

The account abstraction link is explicit in Daisugi. An earlier v0.2 already runs SPHINCS-signed transactions as native Frame Transactions, with per-frame execution and verification alongside the existing ERC-4337 flow, so the same post-quantum signatures slot into the native AA transaction model.

Taken together, the two results move post-quantum signatures toward practicality. Aggregation is what makes large, expensive PQ signatures affordable at network scale, and pairing it with Frame Transactions is how accounts could one day swap to quantum-resistant schemes without the per-signature cost making it impractical.

Start exploring Account Abstraction with Etherspot!

❓Is your dApp ready for Account Abstraction? Check it out here: https://eip1271.io/

Etherspot Website | X | Discord | Telegram | Github | Developer Portal