prompt-guard

To detect prompt injection attacks in user inputs, you can scan the text for malicious patterns, instruction hijacking, and context manipulation using severity tiers for critical and warning signals.

Indirect prompt injection occurs when malicious instructions are hidden in external data sources. It compromises AI safety by manipulating the agent's context and hijacking its intended instructions without direct user input.

You can prevent instruction hijacking from external data sources by employing pattern detection and normalization to produce injection findings and handling guidance before unsafe instructions are trusted by the agent.

Yes, prompt injection detection handles social engineering attempts by identifying direct, indirect, and social engineering-based manipulation patterns across skill content, user inputs, and external data sources.

Yes, you can scan skill content for hidden malicious instructions by analyzing the text for unsafe commands, producing severity-tiered findings, and providing handling guidance to neutralize context manipulation.

Pattern detection for prompt injection may face limitations with novel or heavily obfuscated injection vectors, requiring normalization and severity tiers to effectively filter critical threats from benign warning signals.