ReplayBench T01 - LlamaIndex - Security: leaderboard
Final-state error after replaying an authentic tool result in the wrong task context. T01 attribution experiment: eight models, three separate frameworks, seven separate scenarios and ten valid trials per cell. Lower error is better. This is tool-result misbinding, not injected instructions.
Metric: Final State Error Rate (%). Source: github.com. Saturation forecast: Around January 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 20 | #12 |
| 2 | GLM-5.3 | 20 | #40 |
| 3 | Qwen 3.8 Max | 30 | #31 |
| 4 | GPT-5.5 | 40 | #32 |
| 5 | GPT-5.6 Sol | 40 | #18 |
| 6 | GPT-5.6 Luna | 60 | #69 |
| 7 | DeepSeek V4 Pro | 100 | #99 |
| 8 | Gemini 3.8 Flash | 100 | #21 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=replaybench-t01-llamaindex-security · How It Works · Data refreshed daily, snapshot 2026-10-16.