Unreviewed Components
Skills, plugins, agents, and MCP servers can introduce prompt injection, malicious tools, credential leaks, or backdoors.
AI Agent Security for AI Developers
Start with documented installation paths, review agent components before trust, and evaluate named high-risk actions through the integration available in your development environment.
Security path
AI DevelopersCode / Review / Runtime
01Build
02Review
03Protect
04Ship
Find the Control Point
Coding-agent risk can enter through a trusted dependency or through an action the agent attempts during a session.
Sensitive File and Secret Access
Agent workflows may reach files, credentials, tokens, or other sensitive values.
High-Impact Tool Actions
Shell, tool, and sensitive-write actions can change the development or production environment.
Network and Webhook Activity
Network requests and webhook exfiltration are named runtime action categories that require an observable control path.
Secure Three Moments
Install Through a Documented Path
Use the current Windows, Unix, or npm path described in first-party materials.
Review Components Before Use
Use Deep Scan for documented component and risk categories. Treat the result as review evidence.
Evaluate High-Risk Actions
Use Runtime Guard where a hook, plugin, skill, command, or API path exposes the action for evaluation.
Match the Integration
First-party materials list Claude Code, OpenClaw, Hermes, Codex CLI, Gemini CLI, Cursor, GitHub Copilot, and MCP hosts, but they do not all use the same protection path.
Hooks
Review documented pre/post-tool or native tool-hook paths for the target environment.
Plugin
Use the documented OpenClaw plugin path and verify its current behavior.
Skill or Command
Confirm what the selected skill- or command-based integration can observe and enforce.
API or MCP Host
Verify the request path and keep third-party MCP runtime limitations visible.
Inspect the Implementation
Column 1 titleCLI
Column 1 copyThe public repository documents npm installation and commands including hook, scan, trust, status, and config.
Column 2 titleAPI
Column 2 copyThe API Reference lists runtime analysis, runtime decision, repository scan, package scan, MCP-server scan, and URL scan endpoints.
Column 3 titleData Path
Column 3 copyPublic materials distinguish local handling from cloud-connected sanitized metadata, decisions, policy versions, and audit events.
Verify current authentication, request behavior, data handling, and output before production integration.
Run Developer Tests
Install from the current Quickstart.
Confirm the integration mode.
Scan the components in scope.
Test one expected allow path and one high-risk path.
Inspect the resulting evidence.
Document residual gaps and ownership.
Version and configuration
Test input
Observed decision
Data sent
Actions outside the control path
FAQ
Do all listed integrations provide the same depth?
No. The documented modes differ across hooks, plugins, skills, commands, APIs, and MCP hosts.
Is an API key required for basic local protection?
The public repository says an API key is optional for cloud features. Verify the current setup for the workflow you plan to use.
Which API surfaces are documented?
Runtime analysis, runtime decision, repository scan, package scan, MCP-server scan, and URL scan endpoints are listed.
Can every third-party MCP runtime call be monitored?
No. The public FAQ states that current limitation.
Secure a Coding Workflow
Define the users, decisions, evidence, and ownership required before choosing the implementation path.