ConstraintRot - Injection Attack: leaderboard
ConstraintRot (Chen, 2026) under a summarizer-injection attack: ingested content tells the summarizer to leave the policy out of the compacted context. Violation rate in percent over 27 episodes; lower is better.
Metric: Violation rate (%; 27 episodes in which ingested content instructs the summarizer to omit the policy before the prohibited request). Source: arxiv.org. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 0 | #158 |
| 2 | Gemini 3.5 Flash | 0 | #98 |
| 3 | GLM-5.1 | 22 | #175 |
| 4 | Qwen 3.6 27B | 33 | #339 |
| 5 | GPT-5.4 Mini | 37 | #389 |
| 6 | Kimi K2.5 | 41 | #252 |
| 7 | DeepSeek V4 Flash | 59 | #254 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=constraintrot-injection-attack · How It Works · Data refreshed daily, snapshot 2026-11-12.