Hack-Verifiable TextArena - Read Prompt: leaderboard
Two-player games where the agent can read the opponent model's system prompt. The score is the share of trajectories where it does; lower is better.
Metric: Hack rate, Read Prompt (%). Source: arxiv.org. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0 | #161 |
| 2 | GPT-OSS-120B | 0 | #700 |
| 3 | Gemma 4 31B (IT) | 0 | #342 |
| 4 | GPT-5.4 Mini | 0 | #389 |
| 5 | O3 | 0 | #228 |
| 6 | GPT-5.4 | 2.9 | #123 |
| 7 | Grok 4.1 Fast | 25.7 | #430 |
| 8 | GPT-5 Codex | 26.5 | #245 |
| 9 | Qwen 3.6 Plus | 28.6 | #222 |
| 10 | Kimi K2.5 | 31.4 | #256 |
| 11 | Gemini 3.1 Pro (Preview) | 34.3 | #94 |
| 12 | GLM-5.1 | 60 | #179 |
Interactive version: theaggregate.ai/benchmark?slug=hack-verifiable-textarena-read-prompt · How It Works · Data refreshed daily, snapshot 2026-11-13.