Exfiltration Detection on Single-turn benchmark mixture held-out
94Detection RateAIS
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| AISVariant=full2026.06 | 94 | 1.8 | 16 | |
| AISVariant=CIFT only2026.06 | 88 | 1.5 | 1 | |
| Canary + input filter2026.06 | 82 | 15 | 122 | |
| AISVariant=API-only2026.06 | 71 | 0.8 | 14 | |
| Input filter2026.06 | 60 | 15 | 120 | |
| Output sanitizer2026.06 | 55 | 8 | 50 | |
| Canary text detector2026.06 | 30 | 0 | 2 |