Comprehensive protection
Seven threat classes covered out of the box: prompt injection (attempts to override system instructions), jailbreaks (bypassing safety measures), data exfiltration (extracting secrets or training data), PII exposure, toxicity, unauthorized actions (commands outside permitted scope), and indirect injection (hidden instructions in external content).
Real-time detection
Every check runs in milliseconds with minimal latency overhead, so detection never slows the request path. Full streaming support is maintained.
High accuracy and full visibility
ML-powered models deliver industry-leading precision and recall. Every blocked threat is recorded with detailed analytics and audit logs so you can see exactly what was stopped and why.
Customizable policies
Define what blocks, redacts, or passes through with fine-grained control per policy, per key, and per environment.