OI-Bench - Threat Framing - Attack Success Rate: leaderboard
OI-Bench (Liou et al., 2026) attack success rate (%) for threat framing directives (virus and bomb threats, a zero-score penalty and a penalty-backed override instruction) placed in an appended fifth option: the share of samples whose originally correct answer becomes incorrect once the option is added, whether or not the option itself is chosen. Lower is better.
Metric: Attack Success Rate (%). Source: arxiv.org. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 5.04 | #568 |
| 2 | Grok 4.1 | 6.29 | #473 |
| 3 | Llama 4 Scout | 9.9 | #1787 |
| 4 | GPT-5 Mini | 11.75 | #349 |
| 5 | Qwen 3 8B | 17.43 | #1838 |
| 6 | Qwen 3 235B A22B | 19.06 | #622 |
| 7 | Llama 4 Maverick | 19.18 | #1105 |
| 8 | GPT-5 | 19.68 | #163 |
| 9 | DeepSeek R1 | 23.3 | #467 |
| 10 | DeepSeek V3.2 | 23.99 | #403 |
| 11 | Gemini 2.5 Pro | 39.13 | #262 |
| 12 | Gemini 2.5 Flash Lite | 42.7 | #958 |
Interactive version: theaggregate.ai/benchmark?slug=oi-bench-threat-framing-attack-success-rate · How It Works · Data refreshed daily, snapshot 2026-11-12.