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

Companion repository for the paper submitted to METAVERSE 2026 (Springer LNCS).

Author: Anmol Vats, NYU Tandon School of Engineering (av3938@nyu.edu)

Reproducible capsule: Code Ocean DOI 10.24433/CO.0475286.v1

---

What this repo contains

Metaverse platforms rely on Solana and Algorand smart contracts for NFT ownership, virtual economies, and cross-chain bridges — yet almost no LLM-based security tooling exists for these chains. This repository provides the complete benchmark and pipeline from the paper: 24 vulnerable/patched contract pairs across 8 vulnerability classes, evaluated under 3 prompting strategies with 3 frontier LLMs (GPT-4o, Claude Sonnet 4, Llama-3.3-70B), yielding 216 experimental runs.

---

Repository Layout

Path	Contents
contracts/solana/	Rust/Anchor contracts for V1–V5 (vulnerable and patched)
contracts/algorand/	PyTEAL contracts for V6–V8 (vulnerable and patched)
prompts/	Prompt templates for zero-shot, CoT, and RAG strategies
rag_corpus/	Reference documents used for RAG retrieval
scripts/	Pipeline: RAG index builder, experiment runner, scorer, figure generator
results/raw_outputs/	Raw JSON outputs for all 216 experimental runs
results/scores.csv	Scored results (DR, FPR, EQS, RC) for all 216 runs
figures/	Generated figures (fig1–fig6)
paper/main.tex	LaTeX source (IEEEtran version)
paper/metaverse_main.tex	LaTeX source (LNCS version for METAVERSE 2026)

---

Vulnerability Classes

ID	Chain	Class	Metaverse Impact
V1	Solana	Missing signer check	Unauthorized NFT minting, DAO takeover
V2	Solana	Account confusion (type confusion)	Fake token mints in virtual marketplaces
V3	Solana	Arithmetic overflow on u64	Token vault exploits in play-to-earn economies
V4	Solana	Bump seed canonicalization	PDA signature forgery for program-owned accounts
V5	Solana	Stale account data after CPI	Double-withdrawal in DeFi/metaverse protocols
V6	Algorand	Logic signature abuse	Unauthorized transfers in NFT bridges
V7	Algorand	Group transaction manipulation	Fraudulent injection into atomic NFT settlements
V8	Algorand	Unchecked asset receiver and fee fields	Fund redirection and account takeover

---

Key Results

Strategy	DR (Solana)	DR (Algorand)	FPR (avg)
Zero-shot	96.7%	88.9%	4.2%
Chain-of-thought (CoT)	100%	100%	0.0%
RAG	100%	88.9%	11.1%

CoT achieves 100% detection with 0% false positives across all three models. GPT-4o zero-shot misses account confusion (V2) in 2 of 3 instances. Llama-3.3-70B hallucinates EVM concepts (reentrancy, gas costs) in ~25% of Algorand zero-shot runs.

---

Reproduction

Install dependencies:

pip install -r requirements.txt

Set API keys:

export OPENAI_API_KEY=...
export ANTHROPIC_API_KEY=...
export TOGETHER_API_KEY=...

Build the RAG index:

python scripts/rag_index.py

Dry-run a single contract to verify setup:

python scripts/run_experiments.py \
  --contract contracts/solana/vulnerable/v1_missing_signer_inst1.rs \
  --strategy zero_shot \
  --model gpt-4o \
  --dry-run

Run all 216 experiments:

python scripts/run_experiments.py

Check for any missing runs:

python scripts/run_experiments.py --check-missing

Score outputs:

python scripts/score_outputs.py

Generate all figures:

python scripts/analyze.py

Estimated API cost: under $20 for the full 216-run experiment.

---

Models Evaluated

Model	Provider	Version
GPT-4o	OpenAI	gpt-4o-2024-08-06
Claude Sonnet 4	Anthropic	claude-sonnet-4-20250514
Llama-3.3-70B-Instruct	Meta (via Together AI)	Llama-3.3-70B-Instruct-Turbo

---

Citation

@inproceedings{vats2026metaverse,
  author    = {Vats, Anmol},
  title     = {Securing Metaverse Blockchain Infrastructure: {LLM}-Assisted
               Vulnerability Detection on {Solana} and {Algorand} Smart Contracts},
  booktitle = {Proceedings of METAVERSE 2026 (Springer LNCS)},
  year      = {2026}
}

(BibTeX will be updated with volume/pages after proceedings publication.)