What is Behavry
Behavry is an identity-aware policy enforcement platform for AI agents. It sits between autonomous agents and the systems they interact with — authenticating every action, evaluating it against policy, logging it, and either allowing, modifying, or blocking it.
Think of it as Active Directory + UEBA + policy enforcement, purpose-built for autonomous AI agents.
"Runtime telemetry tells you what happened. Behavry tells you whether it should have."
How it works
Every tool call from every agent flows through the Behavry MCP Proxy:
- Authenticates the agent (JWT RS256, OAuth 2.1)
- Scans the call for DLP violations and inbound injection
- Evaluates the call against OPA policy (allow / deny / escalate)
- Logs the decision with a tamper-proof hash chain
- Monitors behavioral baselines and fires alerts on anomalies
Who it's for
| Persona | Problem Behavry solves |
|---|---|
| CISOs | Auditability, compliance evidence, least-privilege enforcement for AI |
| Security Architects | Policy-as-code for agent behavior, runtime enforcement, SIEM integration |
| Platform Engineers | One proxy to govern all agents — no per-agent instrumentation |
| Compliance Teams | Pre-mapped controls for SOC 2, ISO 27001, NIST AI RMF, EU AI Act |
Key capabilities
| Capability | Description |
|---|---|
| Agent Identity | Every agent gets a cryptographic identity; no anonymous tool calls |
| Policy Engine | OPA with Rego policies — allow, deny, or escalate |
| DLP Scanner | 26 outbound patterns; blocks sensitive data before it leaves |
| Inbound Injection | 7-class scanner for prompt injection in tool results (AOC-1) |
| Requester Identity | Traces which entity instructed each action (AOC-3) |
| Behavioral Monitor | Per-agent baselines, anomaly detection, six-dimension risk score |
| HITL Escalation | Human-in-the-loop approval queue for borderline actions |
| Compliance | Pre-mapped to SOC 2, ISO 27001, NIST AI RMF, EU AI Act, GDPR, HIPAA |
Quick start
git clone https://github.com/behavry/behavry
cd behavry
make dev # starts backend + OPA + TimescaleDB
make setup # provisions dev agent + applies MCP config
Open the dashboard at http://localhost:5173.