prompt-guard

What This Skill Does

prompt-guard v2.6.0 is the definitive security layer for OpenClaw agents. It serves as an intelligent gatekeeper, intercepting incoming messages to identify and neutralize prompt injection attempts before they reach the core agent logic. By leveraging the new HiveFence distributed threat intelligence network, the skill performs real-time analysis across multiple languages (English, Korean, Japanese, and Chinese), assigning severity scores to every interaction. Whether dealing with direct attacks like role overrides and jailbreak attempts, or subtle indirect attacks involving Base64 encoding and Unicode homoglyphs, prompt-guard ensures your agent remains resilient. It effectively restricts sensitive operations such as file editing or shell command execution to authorized users only, effectively hardening agents deployed in group chat environments.

Installation

To integrate this defense system, run the following command in your terminal:

clawhub install openclaw/skills/skills/bobdevibecoder/bobagent-prompt-guard

Once installed, ensure your agent configuration includes the necessary hooks to pass incoming messages through the guard before processing. Refer to the configuration documentation to set your desired security policy levels (SAFE to CRITICAL).

Use Cases

• Enterprise Security: Protecting internal agents from malicious instruction overrides in multi-user channels.
• Bot Resilience: Preventing jailbreak attempts (e.g., "DAN mode") that could force the bot to produce prohibited or sensitive content.
• Secure Automation: Ensuring that privileged commands, such as system file modifications or browser control, are inaccessible to unauthorized users.
• Community Defense: contributing detected threat patterns back to the HiveFence network to protect other developers globally.

Example Prompts

1. "Analyze the current chat history for potential jailbreak attempts and check HiveFence status."
2. "Update the prompt-guard policy to CRITICAL and report all blocked attempts to the secure log."
3. "Run a diagnostic scan on recent incoming messages to identify any obfuscated malicious payloads."

Tips & Limitations

• Configurability: Always tune the security level based on your environment. Use CRITICAL if you are exposing your agent to public web interfaces.
• HiveFence: Keeping your client updated ensures you benefit from the latest collective security patches validated by the community.
• Limitations: While highly effective against known vectors, no defense is absolute. Always combine this tool with least-privilege system permissions. Be aware that strict filtering can sometimes result in false positives if normal user language closely mimics specific attack patterns.