(METAVERSE 2026) Securing Metaverse Blockchain Infrastructure: LLM-Assisted Vulnerability Detection on Solana and Algorand Smart Contracts

Metaverse platforms increasingly rely on blockchain-based smart contracts to govern virtual asset ownership, NFT provenance, in- world economies, and cross-chain token bridges. While Ethereum has received extensive automated security tooling, the two blockchains most actively adopted in metaverse infrastructure, Solana and Algorand, remain almost entirely uncovered by LLM-based audit tools. This paper addresses that gap through a controlled benchmark in which three frontier models (GPT-4o, Claude Sonnet 4, Llama-3.3-70B-Instruct) are evaluated under zero-shot, chain-of-thought (CoT), and retrieval-augmented generation (RAG) prompting across 24 annotated smart contracts spanning eight vulnerability classes on both chains. Each contract is paired with a patched version, enabling false positive rate measurement on known-clean code. From 216 total experimental runs, CoT prompting achieves 100% detection rate with zero false positives across all three models. RAG improves Algorand detection but raises false positive rates; Claude Sonnet 4 under RAG reaches 16.7%. GPT-4o zero-shot misses the subtlest Solana class (account confusion) in two of three instances. Claude Sonnet 4 produces substantially richer vulnerability explanations (EQS 4.24) than GPT-4o (2.88) or Llama-3.3-70B (3.01). Llama-3.3-70B hallucinates EVM-specific concepts into Algorand responses in roughly 25% of zero-shot runs.