API Reference
POST /intercept/prompt
POST
/api/v1/intercept/promptScans an LLM prompt for hardcoded secrets, PII, database credentials, private keys, and other policy violations.
Request
{
"prompt": string // Required. The full prompt text to scan.
}Response
{
"schema_version": "1.1",
"request_id": string,
"allowed": boolean,
"risk_score": number, // 0–100
"risk_rating": string, // "Low" | "Medium" | "High" | "Critical"
"treatment_strategy": string, // "Accept" | "Mitigate" | "Redact" | "Avoid"
"priority": string, // "Critical" | "High" | "Normal"
"priority_reason": string, // only when priority is above risk_rating
"redacted_payload": string, // full prompt, secrets and PII masked (never cut)
"redacted_preview": string, // first 2,000 characters of redacted_payload
"truncated": boolean, // true when redacted_preview was cut
"findings": [{
"rule_name": string,
"match_text": string, // never empty when allowed = false
"description": string,
"severity": string, // this finding alone, same scale as risk_rating
"score": number | null, // 0–1 for AI and plain-language rules
"spans": [{ "start": number, "end": number, "text": string }]
}],
"latency_ms": number, // server processing time, 2 decimals
"latency_us": number, // the same, in microseconds
"cold_start": boolean // true for a one-off slower warm-up request
}latency_msis the time BLAZLE spent on the server. Time measured by your client (for example the Playground's round trip) also includes the network.spansoffsets count characters (Unicode code points) of the payload you sent; span text is masked likeredacted_payload. For the AI detector, which has no exact match, the span is the sentence it most likely reacted to.- Schema 1.1 only adds fields to 1.0, and
latency_msnow has decimals. The same response shape is returned by/intercept/command,/diffand/output.