Vulnerability Mitigation on 792 malicious prompts MCP vulnerabilities 1.0 (test)
0.04ASR (Trial)SPELLSMITH
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SPELLSMITHPrompt Defense Setting=Ident. (Identified vulnerability description), Reflection Stage=Post-reflection2026.07 | 0.04 | 0.13 | |
| SPELLSMITHPrompt Defense Setting=Wrong (Incorrect description), Reflection Stage=Post-reflection2026.07 | 0.17 | 0.51 | |
| SPELLSMITHPrompt Defense Setting=Decl. (Generic declaration), Reflection Stage=Post-reflection2026.07 | 0.67 | 1.52 | |
| SPELLSMITHPrompt Defense Setting=None, Reflection Stage=Post-reflection2026.07 | 2.19 | 3.16 | |
| SPELLSMITHPrompt Defense Setting=Ident. (Identified vulnerability description), Reflection Stage=Pre-reflection2026.07 | 6.69 | 9.09 | |
| SPELLSMITHPrompt Defense Setting=Wrong (Incorrect description), Reflection Stage=Pre-reflection2026.07 | 9.89 | 12.75 | |
| SPELLSMITHPrompt Defense Setting=Decl. (Generic declaration), Reflection Stage=Pre-reflection2026.07 | 21.63 | 28.41 | |
| SPELLSMITHPrompt Defense Setting=None, Reflection Stage=Pre-reflection2026.07 | 56.61 | 63.89 |