Create a project
Name the agent, describe its tools and choose whether policy starts in monitor or enforce mode.
About 2 minutes
Interactive product demonstration · no login
See AgentRiskLayer stop a dangerous AI-agent action.
A fictional support agent receives a hidden instruction inside a customer email and attempts a £2,500 dry-run refund. Follow the complete evidence chain from declared controls to retest and deployment decision.
Safe demonstration: synthetic customer, fake email, dry-run refund tool and no real payment action.
Support agent · scenario overview
MONITORING
A customer support request
Understanding the agent
No action yet
Step 1 of 8
What a real customer actually does
Four practical steps—not a mysterious security dashboard.
AgentRiskLayer connects the agent, policy decision, evidence and remediation in one repeatable workflow.
Connect the Guard API
Send prompts, outputs and proposed tool calls to AgentRiskLayer before your application acts.
One API integrationSet the rules
Allow safe actions, block dangerous patterns and require a server-issued approval bound to the exact sensitive operation.
Versioned policyReview and improve
See blocked events, assign fixes, retest the same attack and retain signed privacy-safe evidence.
Auditable result
In plain English
AgentRiskLayer sits between an AI agent and the action it wants to take.
It checks the request against your policy. A safe action can continue. A dangerous action is blocked. A sensitive action can pause until an authorised person approves the exact customer, amount and expiry.
AI agent→AgentRiskLayer→Tool or API
allow · block · require approval · record evidence
Try it with one real agent
Start free, connect one project and see the first policy decision.
Community includes one project and 10,000 Guard decisions each month. No card is required.