OI-Bench - Instructional Interference - Attack Success Rate: leaderboard
OI-Bench (Liou et al., 2026) attack success rate (%) for instructional interference directives (flattery, self-contradictory instructions, fake chain-of-thought tags and a direct command) 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. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 3.44 | #163 |
| 2 | Claude Haiku 4.5 | 6.2 | #568 |
| 3 | Llama 4 Maverick | 6.3 | #1105 |
| 4 | Llama 4 Scout | 7.55 | #1787 |
| 5 | Grok 4.1 | 8.55 | #473 |
| 6 | GPT-5 Mini | 9.35 | #349 |
| 7 | Gemini 2.5 Pro | 13.33 | #262 |
| 8 | Qwen 3 8B | 18.19 | #1838 |
| 9 | Qwen 3 235B A22B | 18.82 | #622 |
| 10 | DeepSeek V3.2 | 22.13 | #403 |
| 11 | Gemini 2.5 Flash Lite | 24.07 | #958 |
| 12 | DeepSeek R1 | 24.88 | #467 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=oi-bench-instructional-interference-attack-success-rate · How It Works · Data refreshed daily, snapshot 2026-11-12.