NDSS 2025 – PropertyGPT

Why It Matters

Automating property generation cuts expert labor and dramatically improves security assurance for billions‑worth of smart‑contract assets.

Key Takeaways

• LLMs generate verifiable smart‑contract properties automatically
• Retrieval‑augmented prompts improve property relevance
• Iterative compilation feedback ensures syntactic correctness
• Achieved 80% recall against ground‑truth properties
• Detected 26 CVEs and 12 zero‑day bugs

Pulse Analysis

Formal verification has long been the gold standard for proving smart‑contract correctness, yet its adoption stalls because writing comprehensive invariants, pre‑ and post‑conditions demands deep expertise. Traditional tools rely on manual property authoring, a bottleneck that limits coverage and leaves high‑value contracts exposed to subtle bugs. By integrating large language models with a curated property database, PropertyGPT bridges this gap, turning natural‑language specifications into machine‑readable assertions that can be fed directly into provers.

The core innovation lies in a retrieval‑augmented workflow: relevant properties are fetched from a vector store, presented to an LLM for in‑context learning, and then iteratively refined using compilation and static‑analysis signals as an external oracle. A weighted ranking algorithm selects the top‑K candidate properties, which a dedicated prover validates. This loop not only guarantees syntactic correctness but also aligns generated properties with the semantic intent of the original code, achieving an impressive 80% recall compared with expert‑crafted baselines.

For the blockchain ecosystem, PropertyGPT promises a scalable, cost‑effective security layer. Detecting 26 out of 37 known CVEs and surfacing 12 zero‑day flaws demonstrates tangible risk reduction, translating into real‑world bounty earnings. As smart‑contract platforms mature, integrating LLM‑driven verification could become a standard compliance step, accelerating deployment while safeguarding assets worth billions. The approach also hints at broader applications, where AI‑assisted formal methods may soon automate verification across diverse software domains.