CIPHER - Neutral Prompts: leaderboard

Metric: Vulnerability rate (% of 750 generations with at least one vulnerability; 150 neutral prompts x 5 samples; GPT-4o judge, majority of several passes, CIPHER taxonomy). Source: arxiv.org. Saturation forecast: Around 2035. 7 models tracked.

Top models

#ModelScore
1GPT-5.168.8
2CodeLlama-7B-Instruct-hf71.3
3GPT-5.1 Codex73.6
4Gemini 2.5 Flash85.7
5Claude Sonnet 4.590.2

Interactive version: theaggregate.ai/benchmark?slug=cipher-neutral-prompts · How It Works · Data refreshed daily, snapshot 2026-09-25.