DocsIntroduction
Introduction
Verdicter is a runtime security platform for AI agents. It sits between your agent and the real world, evaluating every tool call against your policies before anything is executed.
Whether your agents are sending emails, querying databases, calling external APIs, or modifying files, Verdicter intercepts the action, applies your rules, and returns a decision. A full API call completes in 23ms at the median; the policy decision itself takes 0.02ms.
What Verdicter does
- Evaluates every agent tool call in real time against your policy set
- Returns ALLOW, DENY, or ESCALATE with a reason and risk score
- Logs every decision with full context, payload, and policy trace
- Detects and blocks prompt injection attempts via Shield
- Generates compliance evidence for SOC 2, GDPR, HIPAA, and the EU AI Act
How it works
Every evaluate call passes through three stages: a permission check (is this action listed for this agent?), a policy scan (does any rule match this payload?), and a risk scorer (composite score 0–100). The highest-priority result wins - DENY beats ESCALATE beats ALLOW.
Your Agent Code
verdicter SDK
POST /v1/evaluate
Verdicter Runtime
Permission
Check
Check
→
Policy
Engine
Engine
→
Risk
Score
Score
ALLOW
DENY
ESCALATE
Tool Executes (or doesn't)