Glossary

Key terms used throughout IdenticAPI documentation and API responses.

Verdict
The detector's assessment of the analyzed input. Security detectors return safe, suspicious, or unsafe. Agent Action Guard returns allow, review, or block. Unified Guard normalizes all checks to allow, review, or block and aggregates them into a single decision.
Risk
A severity level accompanying the verdict: low, medium, high, or critical. Maps to verdict — safe/allow → low, suspicious/review → medium, unsafe/block → high.
Confidence
A score from 0 to 1 indicating how strongly the detector matched its patterns or classification. Higher values mean stronger signal. Not all endpoints return confidence.
Finding
A structured detection result with category, reason, optional confidence, and optional character offsets (start/end) into the input text.
Usage unit
The billing unit for IdenticAPI. One unit equals one security check. Unified Guard consumes one unit per check in the request.
Decision
The aggregated outcome from Unified Guard or Agent Action Guard. Uses allow, review, or block. Priority order: block > review > allow.
Prompt injection
An attack where user input attempts to override system instructions, manipulate the model's role, extract hidden prompts, or trigger unauthorized tool execution.
PII (Personally Identifiable Information)
Sensitive personal data such as email addresses, phone numbers, credit card numbers, IBANs, IP addresses, and Social Security numbers detected by pattern matching.
Secret
Credentials and tokens such as private keys, API keys (Stripe, AWS, GitHub, Slack, Google, OpenAI), bearer tokens, and credential key-value pairs.
Redaction
Replacing detected PII and secrets in the input text with placeholders like [EMAIL] or [API_KEY]. Available on the PII & Secrets endpoint via the redact parameter.
Agent Action Guard
A policy engine that evaluates AI agent tool calls (tool name, action, arguments) against configurable rules to allow, review, or block the action before execution.
Policy rule
A condition-action pair in Agent Action Guard. Rules are evaluated by priority (highest first). Condition types include destructive_action, action_type, contains_secret, tool_name, domain_pattern, and data_scope.
Unified Guard
An orchestration endpoint that runs multiple security checks (prompt injection, PII/secrets, output safety, agent action) in a single request and returns a combined decision.
Request ID
A unique identifier (req_...) returned in every response and error. Use it for debugging, support requests, and correlating logs in your dashboard.
Detector version
The version string of the detection engine (currently 1.0.0). Included in responses so you can track behavior changes across releases.