ai-sentinel

What This Skill Does

AI Sentinel is a robust prompt injection firewall designed specifically for OpenClaw agents. It serves as a protective layer that monitors message traffic, tool calls, and tool outputs to identify malicious patterns before they can compromise your agent's integrity. By hooking into the OpenClaw lifecycle, it intercept threats ranging from standard jailbreak attempts to complex indirect prompt injection. Whether you are running a local-only community instance or a globally monitored production agent, AI Sentinel provides the visibility required to maintain a secure AI environment.

Installation

Installation is handled directly through the OpenClaw CLI. First, ensure you are in the root of your OpenClaw project. Run openclaw plugins install ai-sentinel. Once installed, the plugin will require an initialization step to register with your gateway. Use the AskUserQuestion flow to confirm the modification of your openclaw.config.ts. If you opt for the Pro tier, you will be prompted to inject API keys into your .env file, which requires explicit user authorization for file write operations.

Use Cases

1. Enterprise Guardrails: Enforcing strict adherence to internal instructions by detecting jailbreak attempts that try to bypass system prompts.
2. Safe Tool Execution: Preventing indirect prompt injection where malicious data inside tool outputs attempts to manipulate the agent's next action.
3. Compliance & Auditing: Using the Pro dashboard to log security events and analyze the frequency and nature of injection attempts for regulatory reporting.

Example Prompts

1. "Initialize AI Sentinel in Community mode and scan all inbound messages for jailbreak attempts."
2. "Enable the Pro tier for my agent and configure the dashboard reporting for real-time threat analysis."
3. "Check if there are any active prompt injection warnings in the current session log."

Tips & Limitations

• Performance: Community mode operates locally, ensuring zero latency impact. Pro mode adds minimal network overhead for logging.
• False Positives: Like all heuristic scanners, AI Sentinel may occasionally flag complex but benign prompts. Use 'monitor' mode before moving to 'enforce' to tune sensitivity.
• Privacy: Always review the Zetro.ai privacy policy before opting into the Pro tier if you handle sensitive user data, as scan logs will be transmitted externally.