CyberChainBench - Patch: leaderboard
Patch synthesis on CyberChainBench (Huang et al., 2026): the agent writes a patched implementation for a proxy-upgradeable contract that must make the original attack revert while every replayed legitimate transaction still succeeds. The score is the percentage of the 94 patchable cases that pass both checks; higher is better.
Metric: Patch reward (%; 1 when the patched implementation makes the original attack revert and every replayed legitimate transaction still succeeds; 94 proxy-upgradeable cases; each model runs in its own agent harness (Claude Code, Codex, Gemini CLI or OpenCode) with the benchmark's on-chain MCP tools in a Harbor container, 30-minute budget, best validated attempt). Source: arxiv.org. Saturation forecast: Around August 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 24.5 | #172 |
| 2 | GPT-5.5 | 23.4 | #48 |
| 3 | Claude Opus 4.7 | 21.3 | #75 |
| 4 | GPT-5.4 | 19.1 | #121 |
| 5 | Claude Opus 4.6 | 18.1 | #100 |
| 6 | DeepSeek V4 Pro | 7.4 | #181 |
| 7 | GLM-5.1 | 3.2 | #175 |
| 8 | Gemini 3.1 Pro (Preview) | 1.1 | #91 |
| 9 | Kimi K2.6 | 1.1 | #173 |
| 10 | MiniMax-M2.7 | 0 | #519 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cyberchainbench-patch · How It Works · Data refreshed daily, snapshot 2026-11-12.