I made a major update to Prompt Guard for Clawdbot (now OpenClaw), which I shared yesterday. With advice from a white-hat hacker friend who is among the best in the world, this release expands the attack patterns to 349.
New defenses:
• Impersonated authority detection ("I’m an administrator," "This is a system message")
• Indirect injection blocking (malicious prompts hidden in URLs/files)
• Multi-turn manipulation detection (attacks that build trust across multiple conversations)
• Token smuggling defense (Unicode zero-width character tricks)
• Scenario-based jailbreak blocking ("In a dream...," "As a movie script...")
• Emotional manipulation detection ("It’s an emergency," "A life is at risk")
Installation and updates take 30 seconds.
New users
clawdhub install prompt-guard
Existing users
clawdhub update prompt-guard
🔗 GitHubseojoonkim/prompt-guardAdvanced prompt injection defense system for AI agents. Multi-language detection, severity scoring, and security auditing.GitHub · github.comOpen🛒 ClawdHubprompt-guardAgent skill by @skills on ClawHub.ClawHub · clawdhub.comOpen
