Ethereum co-founder Vitalik Buterin has presented a demonstration prototype of an anonymous message board for the L2 network Aztec. The project's source code has already been published on GitHub. This is not just another experiment, but a practical implementation of a concept Buterin first voiced back in 2022 — creating a platform where user anonymity is combined with effective content moderation.
The prototype, which the developer called a "toy version," was created in vibe-coding mode, where code generation largely falls on the shoulders of AI agents. Despite the early stage, the system already demonstrates non-trivial capabilities. A user deposits ETH through a smart contract on the Ethereum mainnet, gaining the right to publish messages on the Aztec network. Funds can be withdrawn back to L1, with the address and deposit amount remaining public, as does the recipient's wallet upon withdrawal. However, the messages themselves do not contain the sender's address and are not tied to the original deposit.
The Economics of Anonymity and Moderation
The anonymity set is formed from all users who have deposited funds but have not yet withdrawn them. The frequency of publications directly depends on the deposit size. With a base delay of 3600 seconds (one hour), the following parameters apply:
- 0.001 ETH — one message per hour;
- 0.005 ETH — five messages per hour;
- 0.01 ETH — ten messages per hour.
Unused limits can be accumulated, and a repeat deposit does not bypass the restriction — a new initial delay is set for it.
A key element of the system is moderation. When deploying the contract, a moderator address is assigned (the censor role). The moderator can mark messages as violating policy, after which they are hidden from the main feed, but the user can view them through a separate warning. For a violation, the author receives an additional delay, the size of which depends on a pre-set coefficient. Interestingly, moderator rights can be transferred to another address, including a multi-signature wallet or a DAO, but the protocol itself does not distribute authority automatically.
AI on Guard
The policy text is stored on-chain and is limited to 1488 bytes. For automatic checking, a local daemon with a large language model is provided, running through llama.cpp. By default, Qwen3.5-2B is used in a compressed format (~1.4 GB). The model runs locally and does not access an external API, which is critical for a decentralized architecture. The daemon monitors new publications, compares them against the policy, and sends a flag to the smart contract.
The repository is in an early stage and does not contain any ready releases. A known issue exists with the update_portal() function, which can be called by any user specifying a spoofed address, potentially locking the contract. Developers will need to strengthen zero-knowledge proofs and address other limitations.
My analysis: This project is an important step towards the practical implementation of private yet moderatable decentralized platforms. The combination of zero-knowledge proofs (via Aztec) and local AI for moderation could become a prototype for future social networks and forums where free speech does not contradict security. However, the discovered vulnerability reminds us that even brilliant concepts require thorough auditing before deployment in real-world products.