MemPoison - Poisoned Task: leaderboard
Agent memory-poisoning benchmark (arXiv 2607.14651); behavioral corruption rate is the share of trigger tasks, run after poisoned records were written to the agent's memory, whose answer matches the attacker's target (lower is better).
Metric: Behavioral Corruption Rate (%). Source: arxiv.org. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 54.67 | #343 |
| 2 | DeepSeek-V2-Lite | 56.96 | #1360 |
| 3 | Llama 3.1 8B | 61.41 | #1151 |
| 4 | Qwen 3 32B | 61.44 | #443 |
| 5 | Qwen 3 8B | 63.59 | #674 |
| 6 | Qwen 2.5 14B | 64.2 | #725 |
| 7 | GLM-4 32B | 65.1 | #459 |
| 8 | DeepSeek V3 | 65.44 | #321 |
| 9 | Gemini 3 Flash | 65.78 | #99 |
| 10 | GPT-5 | 66.87 | #98 |
Interactive version: theaggregate.ai/benchmark?slug=mempoison-poisoned-task · How It Works · Data refreshed daily, snapshot 2026-10-20.