Cross-modal attack detection on SkillMutator GPT-4o-mini adversarial-oracle evaluable subset (n=48)
89.6Detection RateGPT-5.4
Evaluation results
| Method | Links | |
|---|---|---|
| GPT-5.4Scanner Type=Proprietary LLM2026.06 | 89.6 | |
| GPT-5.4-miniScanner Type=Proprietary LLM2026.06 | 79.2 | |
| PIGuardScanner Type=Prompt-Injection Detector2026.06 | 39.6 | |
| GPT-4o-miniScanner Type=Proprietary LLM2026.06 | 35.4 | |
| Snyk Agent ScanScanner Type=Rule-based / Commercial2026.06 | 16.7 | |
| DataSentinelScanner Type=Prompt-Injection Detector2026.06 | 4.2 | |
| skill-security-scanScanner Type=Rule-based / Commercial2026.06 | 2.1 | |
| LLM-GuardScanner Type=Prompt-Injection Detector2026.06 | 2.1 | |
| SkillScanScanner Type=Rule-based / Commercial2026.06 | 0 |