lieutenant

What This Skill Does

Lieutenant serves as the primary security layer for the OpenClaw ecosystem, acting as a specialized firewall for AI agents. It is designed to inspect incoming and outgoing communications, including messages, agent identity cards, and cross-agent traffic, to detect malicious intent. The skill identifies over 65 distinct threat patterns ranging from classic prompt injection and jailbreaking techniques to more complex, modern threats like data exfiltration and conversation history poisoning. By integrating Lieutenant into your workflow, you create a robust barrier that prevents unauthorized access, safeguards sensitive credentials, and ensures that the agents you interact with adhere to established safety protocols.

Installation

To add the Lieutenant security suite to your environment, use the command: clawhub install openclaw/skills/skills/jd-delatorre/lieutenant. Once installed, ensure you have the necessary environment variables set in your local configuration, specifically the OpenAI API key if you require semantic analysis for sophisticated evasion detection, or the TrustAgents API key for real-time reputation monitoring and crowdsourced threat intelligence. You can further customize behavior by setting LIEUTENANT_STRICT=true to block any message containing a potential threat.

Use Cases

Lieutenant is essential for any high-stakes AI automation project. Use it when:

• Communicating with untrusted or external agents to ensure they aren't attempting to inject malicious code or manipulate your system prompts.
• Protecting sensitive internal databases by scanning all outgoing data streams for PII or credential leakage.
• Verifying the authenticity of agent cards through the .well-known/agent.json standard, ensuring that the services you connect to are legitimate.
• Automating secure Agent-to-Agent (A2A) communications by utilizing the LieutenantInterceptor middleware to automatically vet every exchange.

Example Prompts

1. "Lieutenant, scan this recent message from the support agent for prompt injection or data exfiltration attempts: [Paste Message Content]"
2. "Verify the authenticity of this remote agent card to ensure it meets our security standards: https://agent.example.com/.well-known/agent.json"
3. "Run a diagnostic scan on the current conversation history to check for context manipulation or hidden commands."

Tips & Limitations

To maximize the effectiveness of Lieutenant, always enable semantic mode when dealing with complex, multi-step interactions, as simple pattern matching may miss cleverly paraphrased jailbreak attempts. While the tool is powerful, it is not a silver bullet; always pair it with the Principle of Least Privilege regarding your agents' permissions. Note that enabling advanced detection features via the TrustAgents API or OpenAI requires active external service subscriptions. Keep your environment variables updated regularly to ensure you have the latest threat pattern definitions.