CarryoverBench - Under-Attack Utility: leaderboard
CarryoverBench:nanobot persistent-agent runtime,101attack variants and12benign tasks;temp0,sharedprompt/tools. Instant single-shot and Autonomous/Latent any-fire over5followups; utility task-weighted artifact rubric. Qwen3-235B-A22B main run disables thinking; other requested efforts unspecified. Defense/runtime ablations are separate experiments. Rate (%).
Metric: Rate (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 94.2 | #60 |
| 2 | GLM-5 | 93.1 | #140 |
| 3 | Qwen 3 235B A22B (Non-reasoning) | 91.6 | #315 (Qwen 3 235B A22B) |
| 4 | GPT-5.4 | 91.4 | #82 |
| 5 | Claude Sonnet 4.6 | 87.9 | #92 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=carryoverbench-under-attack-utility · How It Works · Data refreshed daily, snapshot 2026-10-22.