Aztec Says an Internal AI-Assisted Review Found a Flaw That Could Make Invalid Proofs Verify

Aztec, the privacy-focused Ethereum layer-2 network, says core contributors found a critical vulnerability in the proving system behind its Alpha V5 release on July 27, using internal auditing assisted by artificial intelligence. In the team's own description of the flaw, an attacker may be able to construct a proof that passes verification for a transaction the network should reject. Contributors say reviewers have identified no other high-severity or critical issues in the current alpha, and that the repair is planned for the next version rather than an emergency patch.

This is the second time Aztec has run this sequence. The team disclosed a critical proving-system vulnerability in Alpha V4 earlier this year, kept the technical details private until the successor release, and told users to withdraw funds before the V5 upgrade vote made the older bug public. That pattern is now standard practice for zero-knowledge systems, where publishing a soundness bug before the fix ships hands an attacker the ability to mint value from nothing.

A soundness failure is a different risk from a smart-contract bug. When a rollup's proof can be forged, re-execution by honest validators does not catch it, because the network was designed to trust the proof instead of repeating the work. The affected component sits below every application built on top of it.

The cryptography one layer down is also under active dispute. On Ethereum Research, a post titled "Poseidon2b is secure!" answers an earlier claim that Poseidon is not safe for Ethereum, arguing that no hash family can be judged secure or insecure without fixing the field, width, capacity, round schedule, matrices and attack model. The pressure is real: researchers at ETH Zurich published round-skipping algebraic attacks that cut Gröbner-basis preimage estimates for some Poseidon2 instances below the 2^128 level that designers target, and the Ethereum Foundation has signalled a preference for the long-studied SHA and BLAKE3 families for protocol use.

Both developments reflect the same underlying condition. Trust in zero-knowledge systems now rests on a small number of proving stacks and arithmetization-friendly hash functions, reviewed by a small number of specialists, and increasingly by automated tools.