Cross-modal attack detection on SkillMutator GPT-5.4-mini adversarial-oracle evaluable subset (n=63)
88.9Detection RateGPT-5.4
Evaluation Results
| Method | Links | |
|---|---|---|
| GPT-5.4Scanner Type=Proprietary LLM2026.06 | 88.9 | |
| GPT-5.4-miniScanner Type=Proprietary LLM2026.06 | 71.4 | |
| DataSentinelScanner Type=Prompt-Injection Detector2026.06 | 11.1 | |
| Snyk Agent ScanScanner Type=Rule-based / Commercial2026.06 | 9.5 | |
| GPT-4o-miniScanner Type=Proprietary LLM2026.06 | 9.5 | |
| skill-security-scanScanner Type=Rule-based / Commercial2026.06 | 6.4 | |
| PIGuardScanner Type=Prompt-Injection Detector2026.06 | 4.8 | |
| SkillScanScanner Type=Rule-based / Commercial2026.06 | 0 | |
| LLM-GuardScanner Type=Prompt-Injection Detector2026.06 | 0 |