Pick what the text is, paste it or click an example, and press Check. The first check downloads the model once (it is then cached on this device). In a browser tab the first check takes a few seconds and later ones about a second; the same model in Python on a server CPU takes about 44 ms per question.
The same check in your code
Same model, same questions, on any CPU: pip install decima-ai.
Why another guard
Most open guards were trained on English. On ordinary Arabic and Persian messages they raise false alarms on most of them, so real customers get blocked. Same frozen public test for both models, int8, CPU:
| decima-guard | ProtectAI prompt-injection-v2 | |
|---|---|---|
| Ordinary Arabic messages let through (589) | 99.7% | 25.1% |
| Ordinary Persian messages let through (205) | 100% | 20.5% |
| Prompt-injection test, 6,532 rows: accuracy | 0.915 | 0.784 |
| False alarms on 4,242 harmless messages | 124 | 979 |
Weak spot: it misses about 1 in 6 real attacks translated into Arabic dialects, and about 1 in 3 translated into Persian. Keep hard rules first and use the guard as one layer. Every test row and prediction is public in decima-guard-bench.