Background: DeFi's Unverified Vulnerability Problem
DeFi has suffered catastrophic losses in recent years: Poly Network was hacked for $611M (August 2021), Ronin Network for $625M (March 2022), and BNB Chain for $570M (October 2022). But a deeper problem lurks: most disclosed vulnerabilities are never verified as actually exploitable. Researchers flag suspicious patterns, developers can't tell real threats from false alarms, and attackers quietly confirm every "maybe."
On May 4, 2026, Ruichao Liang and seven co-authors published a paper on EvoPoC, an AI system that:
- Read 2,573 smart contracts and automatically found vulnerabilities
- Automatically generated attack code and verified it could actually profit
- Achieved a 96.6% exploit success rate
- Reproduced 85 real-world historical DeFi attacks
- Discovered 16 previously unknown zero-day vulnerabilities
- Helped protect $70.6M in assets
- Layer 1 — Concrete patterns: code snippets, historical exploits (e.g., Poly Network's cross-chain verification bypass, Beanstalk's governance flash-loan attack)
- Layer 2 — Exploit primitives: reentrancy, oracle manipulation, access control bypass, integer overflow
- Layer 3 — Protocol semantics: AMMs, flash loans, governance, yield farming
- Title: EvoPoC: Automated Exploit Synthesis for DeFi Smart Contracts via Hierarchical Knowledge Graphs
- Authors: Ruichao Liang, Jing Chen, Xianglong Li, Huangpeng Gu, Yebo Feng, Yue Xue, Cong Wu, Yang Liu
- arXiv: 2605.02868
- Published: May 4, 2026
- Categories: cs.CR, cs.SE
Core Insight: Exploit Synthesis Is Structured Reasoning
The paper's key claim: exploit synthesis is not a code generation task but a structured reasoning problem. Before generating attacks, the AI must understand:
1. Protocol semantics — what the contract does (lending, trading, yield farming) 2. Failure root causes — the nature of the bug (reentrancy, price manipulation, access control bypass) 3. Attack primitives — which known techniques can be composed (flash loans, sandwich attacks)
Hierarchical Knowledge Graph (HKG)
EvoPoC organizes this knowledge in three layers:
LLMs use the HKG for multi-hop reasoning: from concrete code anomalies → matching primitives → semantic understanding → synthesized exploit.
Two-Stage Verification
Generating code is easy; proving it works and profits is hard. EvoPoC verifies every exploit twice:
1. Logical verification: an SMT solver checks whether an execution path from the initial state to the completed attack actually exists. 2. Economic verification: asset-level state simulation tracks token transfers, liquidity pool updates, and price changes to confirm the attack yields profit > 0.
Only exploits passing both stages count as valid proofs of concept.
Results
Detection (2,573 contracts, 72 projects):
| Metric | EvoPoC | Traditional fuzzers | LLM generator A1 | |:---|:---:|:---:|:---:| | Recall | 98% | ~60% | ~75% | | F1-Score | 0.90 | ~0.55 | ~0.65 |
Exploitation:
| Metric | EvoPoC | SOTA fuzzers | LLM generator A1 | |:---|:---:|:---:|:---:| | Exploit success rate | 96.6% | ~19% | ~48% | | Historical attacks reproduced | 85 / 88 | ~30 | ~45 | | Recoverable value | baseline | 1/300× | 1/8.5× |
Real-world impact: $116.2M in assets recovered, 16 zero-days disclosed, $70.6M protected, $2,900 in bug bounties.
The Uncomfortable Truth
The paper's techniques are public and reproducible. The attack-defense asymmetry is being broken in *both* directions: defenders can audit contracts before attackers strike, but attackers could deploy similar systems to automatically scan the entire DeFi ecosystem. EvoPoC found 16 zero-days in projects that had already passed manual audits—evidence that DeFi complexity now exceeds the limits of human review. The future is human + AI collaboration, but the transition period may be dangerous if attackers adopt these tools first.