Read this article on bsc.news A newly published cryptography paper proposes a collision attack on LtHash, the lattice-based hash function that @SeiNetwork uses to produce fingerprints of its
A newly published cryptography paper proposes a collision attack on LtHash, the lattice-based hash function that @SeiNetwork uses to produce fingerprints of its blockchain state. The attack is theoretical, and the resources it demands are far beyond what is practical today.
The paper, titled "Two-Bit Lifting for Ternary SIS: Polynomial-Time Collision Attacks on LtHash" (ePrint 2026/2083), is by Rui Ding, Xiaorui Gong, Hao Jiang and Lili Tang. According to the Cryptology ePrint Archive, LtHash was introduced by Bellare and Micciancio at EUROCRYPT 1997. The archive lists the paper as received on 18 September 2026 and approved four days later. ePrint papers are preprints and are not the same as peer-reviewed publications.
The authors say that for LtHash16, with parameters (N,d) of (1024,16), their attack requires approximately 2^81 hash queries. They note that the same configuration is used at Facebook and in Solana. The abstract also describes a method that cuts the number of queries needed compared with a more basic approach.
Collision attacks of this kind rely on finding two different sets of inputs that produce the same fingerprint. The work involved is not only computational. An attacker would also have to generate and store an enormous number of fingerprints, which is why the scale of the requirement matters. By the figures in the source analysis, reaching the required database size would mean growing Sei's current state by a factor of about 2 trillion. Even at a rate of one billion new fingerprints per second, that would take more than 130 million years.
Sei is moving away from Merkle trees for its EVM state. In a post on its Eidos storage upgrade, the network says a lattice hash maintains a running fingerprint of the entire state. The fingerprint updates in constant time per change, without cascading recomputation. Its replacement store, FlatKV, keeps state in a flat key-value layout where one change is one write.
The rollout is phased. Crypto.news reported that the first Eidos components reached mainnet through the v6.6 upgrade, with FlatKV and LtHash planned for later stages. Sei's documentation also describes LtHash as the basis of state commitments in the planned Giga upgrade, where it would replace the Merkle state root on the hot path.
The paper's findings have also been weighed elsewhere. A security note in a Solana improvement proposal on its accounts lattice hash describes the attack as having a query cost below 2^128 but remaining impractical. It cites upper bounds of 2^101 binary operations and about 2^88.45 bits of memory. These parameters were chosen to target 128-bit collision security, so the result is notable for academic reasons even though it does not translate into a working exploit.
For now, the research is best read as a refinement of the theoretical security margin of LtHash rather than an active threat to Sei or other networks using it. Developers and researchers may still look closely at parameter choices as the technique is reviewed.
Cryptology ePrint Archive: Two-Bit Lifting for Ternary SIS, Paper 2026/2083
Solana Improvement Documents: Accounts Lattice Hash (SIMD-0215)
