MCP-SafetyBench - Task Success: leaderboard
The share of the same 245 attacked MCP tasks in which the agent still completes the user's task. It is the utility side of the benchmark: the paper reports that models completing more tasks also tend to be attacked more successfully.
Metric: Task success rate under attack (%, all 245 tasks; utility control). Source: arxiv.org. Saturation forecast: Around February 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O4 Mini | 21.22 | #353 |
| 2 | Gemini 2.5 Pro | 20.41 | #266 |
| 3 | DeepSeek V3.1 | 19.59 | #497 |
| 4 | GLM-4.5 | 18.37 | #526 |
| 5 | GPT-5 | 15.92 | #167 |
| 6 | Grok 4 | 15.92 | #307 |
| 7 | Gemini 2.5 Flash | 15.1 | #453 |
| 8 | Claude 3.7 Sonnet | 15.1 | #458 |
| 9 | Claude Sonnet 4 | 10.2 | #400 |
| 10 | GPT-4.1 | 9.8 | #495 |
| 11 | GPT-4o | 8.98 | #741 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mcp-safetybench-task-success · How It Works · Data refreshed daily, snapshot 2026-11-13.