System Prompt Exfiltration on Cursor Agent Environment
100Pseudo-RecallToolLeak
Evaluation Results
| Method | Links | |
|---|---|---|
| ToolLeakBackend LLM=grok-42025.09 | 100 | |
| ToolLeakBackend LLM=sonnet-4.52025.09 | 100 | |
| ToolLeakBackend LLM=composer-12025.09 | 100 | |
| ToolLeakBackend LLM=sonnet-42025.09 | 99 | |
| Completion-RememberBackend LLM=gemini-2.5-pro2025.09 | 78 | |
| NaiveBackend LLM=gemini-2.5-pro2025.09 | 68 | |
| IgnoreBackend LLM=gemini-2.5-pro2025.09 | 66 | |
| CompletionBackend LLM=gemini-2.5-pro2025.09 | 45 | |
| Remember-the-StartBackend LLM=sonnet-42025.09 | 23 | |
| NaiveBackend LLM=grok-42025.09 | 0 | |
| NaiveBackend LLM=sonnet-42025.09 | 0 | |
| NaiveBackend LLM=sonnet-4.52025.09 | 0 | |
| NaiveBackend LLM=composer-12025.09 | 0 | |
| ToolLeakBackend LLM=gemini-2.5-pro2025.09 | 0 |