Prompt Exfiltration on Claude Code coding agent prompts
89.1Similarity ScoreToolLeak
Evaluation results
| Method | Links | ||
|---|---|---|---|
| ToolLeak2025.09 | 89.1 | 64.7 | |
| Completion-Remember2025.09 | 55.2 | 65.5 | |
| Naive2025.09 | 54.7 | 96.2 | |
| Ignore-Remember2025.09 | 52.9 | 81.7 | |
| Remember-the-Start2025.09 | 52.4 | 65.9 | |
| Ignore-Completion-Remember2025.09 | 51.3 | 82.2 | |
| Ignore2025.09 | 50.5 | 98.4 | |
| Completion2025.09 | 50.2 | 97.9 | |
| Ignore-Completion2025.09 | 49 | 97.4 | |
| PLeak2025.09 | 21.6 | 97.8 |