Taktile PIBench - Untrusted-Content Tagging: leaderboard
The same underwriting injections, but every untrusted region of the prompt is wrapped in boundary tags and the system prompt tells the model to treat tagged content as data; the number is the share of rollouts defended. It measures how well a model uses that explicit marking, which the headline board withholds.
Metric: Defense success rate (%; over 220 rollouts per model, higher is better). Source: labs.taktile.com. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 100 | #94 |
| 2 | GPT-5 | 100 | #167 |
| 3 | Claude Sonnet 4.5 | 100 | #250 |
| 4 | Claude Sonnet 4.6 | 100 | #161 |
| 5 | GPT-5.2 | 100 | #176 |
| 6 | Claude Opus 4.6 | 100 | #102 |
| 7 | Claude Opus 4.7 | 100 | #79 |
| 8 | GPT-5.1 | 100 | #229 |
| 9 | GPT-5.2 Pro | 100 | #124 |
| 10 | GPT-5 Mini | 98.2 | #348 |
| 11 | Gemini 3 Flash | 97.3 | #164 |
| 12 | Gemini 2.5 Flash Lite | 96.4 | #958 |
| 13 | Gemini 2.5 Pro | 94.5 | #266 |
| 14 | Claude Haiku 4.5 | 94.5 | #568 |
| 15 | Gemini 2.5 Flash | 91.8 | #453 |
Interactive version: theaggregate.ai/benchmark?slug=taktile-pibench-untrusted-content-tagging · How It Works · Data refreshed daily, snapshot 2026-11-13.