little-canary

Detects prompt injection by its effect on a sacrificial canary model, not just pattern matching: untrusted input hits a powerless model first, a behavioral check reads the residue, and it returns block, flag, or pass before your primary model acts. Inbound preflight sensor, not a guarantee.

Provenienza della descrizione: Fonte originale

Misurato: 24 ott 2026 · 43 osservazioni