ARAgentRiskLayerAI agent security
Live protection, explained simply

A safety checkpoint before your agent acts.

Your agent proposes an action. AgentRiskLayer checks it against your rules. Safe actions continue, unsafe actions stop, and sensitive actions wait for approval tied to the exact operation.

Try the safe exampleWatch it work first
What happens in one request

Four decisions the customer can understand.

1

The agent proposes something

A prompt, model output or tool call reaches the security boundary before execution.

2

The current rules are checked

The project policy evaluates content, tool, destination, path, environment and required approval.

3

The action is controlled

The decision is allow, deny or require an exact server-issued approval. Caller claims are not trusted.

4

Privacy-safe evidence is kept

The event records policy identity, rule IDs, digests and timing—not raw prompts or tool arguments.

Preserved technical depth

Choose the integration boundary that fits the system.

A beginner can use the built-in example. A developer can then open the exact integration details.

Hosted

Guard API

Scoped keys, idempotent request IDs, immediate revocation, exact-action approvals and atomic single-use consumption.

Private environment

Customer-operated gateway

Route tool traffic through a local enforcement point when runtime content must remain in the customer environment.

Content

Prompt and output screening

Detect instruction override, hidden extraction, secret leakage, encoded attacks and unsafe network instructions.

Actions

Tool-call enforcement

Allow or deny tools, hosts, paths and material actions using a versioned project policy.

Developer reference

Hosted Guard request

Place this check before the model output or proposed tool call reaches a real system.

Open full quick start
curl -sS https://agentrisklayer.com/v1/guard \
  -H "Authorization: Bearer arl_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "request_id": "agent-step-123",
    "input": "Customer request and retrieved context",
    "tool_call": {
      "name": "crm.read",
      "arguments": {"customer_id": "cust_123"},
      "context": {"environment": "production"}
    }
  }'

A repeated request ID returns the original decision. Approval-required actions use a server-issued approval token; caller approval booleans are ignored.

Private deployment option

Customer-operated gateway

Keep runtime content inside your network and route protected tool traffic through the local gateway.

node agent-risk-runtime.mjs \
  --policy runtime-policy.json \
  --upstream http://127.0.0.1:3000 \
  --port 8787 \
  --audit runtime-audit.jsonl
Download runtime gatewayDownload example policy

The downloadable gateway does not contain the hosted PostgreSQL approval ledger. Local approval-required actions therefore fail closed rather than trusting caller assertions.

Security boundary

Protection only works where traffic is routed through it.

Using MCP servers or MCP-enabled agents? Review the MCP server risk assessment to check tool trust, permissions, secrets and action validation before deployment.