Runtime guard for agents

Scan agent inputs and tool calls before they turn into action.

Armorer Guard is the local inspection layer for agent runtimes. It helps catch prompt injection, credential leakage, exfiltration attempts, and dangerous tool calls at the boundary where text crosses into execution.

Inspect a tool call

$ cat tool-call.json | armorer-guard inspect-json

{

"policy_action": "dangerous_tool_call",

"confidence": "high",

"sanitized": true

}

$armorer guard openclaw --watch

Where Guard fits

Put policy at the execution boundary.

Prompt-only defenses are brittle. Guard is designed for the places where agent text can influence commands, files, tools, logs, memory, and outbound messages.

Inspect untrusted text

Scan web pages, documents, tickets, chat messages, retrieved snippets, and model output before they enter agent context.

Gate risky actions

Evaluate tool-call arguments before shell, filesystem, browser, SaaS, or MCP actions can become system changes.

Return structured policy

Use JSON results with sanitized text, reasons, confidence, and policy labels inside an orchestrator or control plane.

Detection coverage

Signals Guard can flag.

These checks are useful by themselves, and stronger when paired with Armorer runtime controls for sandboxing, credentials, approvals, and audit trails.

prompt injection
system prompt extraction
data exfiltration
sensitive data requests
safety bypasses
destructive commands
credential disclosure
dangerous tool calls