System Prompt Exfiltration on Copilot Agent Environment
99Pseudo-RecallToolLeak
Evaluation Results
| Method | Links | |
|---|---|---|
| ToolLeakBackend LLM=sonnet-42025.09 | 99 | |
| ToolLeakBackend LLM=sonnet-4.52025.09 | 98 | |
| Ignore-CompletionBackend LLM=gemini-2.5-pro2025.09 | 87 | |
| Ignore-RememberBackend LLM=gemini-2.5-pro2025.09 | 87 | |
| Remember-the-StartBackend LLM=sonnet-4.52025.09 | 42 | |
| Remember-the-StartBackend LLM=sonnet-42025.09 | 30 | |
| ToolLeakBackend LLM=gemini-2.5-pro2025.09 | 13 |