CIPHER: leaderboard
Cryptographic code-security benchmark (CIPHER, Manolov, Gao, Shukla, Chou and Lagasse, 2026): 150 Python cryptography tasks, each posed as an insecure, a neutral and a secure prompt, with five samples per prompt (2,250 generations). A GPT-4o judge, run several times with a majority vote, labels each generation against a taxonomy of 58 cryptographic vulnerability types; an omitted security-critical step counts as a vulnerability. The score is the percentage of generations with at least one vulnerability; lower is better.
Metric: Vulnerability rate (% of 2,250 generations with at least one vulnerability; 450 prompts x 5 samples; GPT-4o judge, majority of several passes, CIPHER taxonomy). Source: arxiv.org. Saturation forecast: Around 2034. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 70 | #225 |
| 2 | GPT-5.1 Codex | 70.9 | #279 |
| 3 | CodeLlama-7B-Instruct-hf | 76.2 | #9444 |
| 4 | Gemini 2.5 Flash | 80.7 | #452 |
| 5 | Claude Sonnet 4.5 | 89.1 | #244 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cipher · How It Works · Data refreshed daily, snapshot 2026-11-12.