MOSAIC-Bench: leaderboard
Measures whether coding agents implement exploitable code when a vulnerability is split across three routine engineering tickets: 199 chains over 10 containerized web-app substrates and 31 CWE classes, one fresh agent session per ticket, and a deterministic proof-of-concept exploit oracle on the deployed result. Strict attack success rate in percent (exploitable share of implementations that still build), lower is better; Claude models ran in Claude Code, GPT models in Codex and others in OpenCode. Steinberg and Gal (University of Haifa), 2026.
Metric: Strict attack success rate: exploitable share of built three-ticket implementations (%, lower is better). Source: arxiv.org. Saturation forecast: Around December 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 53.3 | #100 |
| 2 | Claude Sonnet 4.6 | 58.8 | #158 |
| 3 | Gemini 3.1 Pro (Preview) | 65.1 | #91 |
| 4 | MiniMax-M2.7 | 65.3 | #519 |
| 5 | GLM-5 | 69 | #256 |
| 6 | Kimi K2.6 | 70.1 | #173 |
| 7 | Gemini 3 Flash | 70.4 | #159 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mosaic-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.