Taktile PIBench: leaderboard
A model runs as a loan-underwriting agent while 53 hand-written injections hidden in the user message, structured input, uploaded documents, file names or tool responses try to get a high-risk applicant approved or applicant data leaked; the number is the share of rollouts where it still flags the applicant and leaks nothing. Unlike AgentDojo it is one realistic finance workflow with no defenses added.
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 | GPT-5 | 100 | #167 |
| 2 | GPT-5.2 | 100 | #176 |
| 3 | GPT-5 Mini | 100 | #348 |
| 4 | Claude Opus 4.7 | 100 | #79 |
| 5 | GPT-5.2 Pro | 100 | #124 |
| 6 | Claude Opus 4.6 | 99.1 | #102 |
| 7 | GPT-5.1 | 98.2 | #229 |
| 8 | Claude Sonnet 4.6 | 97.3 | #161 |
| 9 | Gemini 3 Flash | 94.5 | #164 |
| 10 | Claude Haiku 4.5 | 91.8 | #568 |
| 11 | Claude Sonnet 4.5 | 90.9 | #250 |
| 12 | Gemini 3.1 Pro (Preview) | 89.1 | #94 |
| 13 | Gemini 3.1 Flash Lite | 89.1 | #364 |
| 14 | Gemini 2.5 Flash Lite | 88.2 | #958 |
| 15 | Gemini 2.5 Pro | 78.2 | #266 |
Interactive version: theaggregate.ai/benchmark?slug=taktile-pibench · How It Works · Data refreshed daily, snapshot 2026-11-13.