prompt-guard

Prompt injection attacks are detected by validating user inputs and third-party data using a pre-trained 86M parameter sequence classifier. This identifies and blocks malicious instructions embedded within prompts before they reach the LLM.

Preventing jailbreak attempts involves scanning incoming user messages for direct overrides of safety guidelines using a dedicated classifier. This flags malicious inputs and ensures the LLM maintains its intended behavior and safety alignment.

Yes, you need the transformers and torch libraries to run prompt injection detection. These dependencies are required to load the pre-trained 86M parameter classifier and perform sequence classification inference.

Yes, you can use a pre-trained classifier to filter malicious instructions in third-party data. The model validates external inputs for embedded malicious instructions, ensuring secure LLM interactions across all data sources.

Input validation for LLM security catches both direct and indirect prompt injections. It identifies direct jailbreak attempts overriding safety guidelines and flags malicious instructions hidden within third-party data.

The 86M parameter classifier for jailbreak detection requires torch and transformers dependencies and focuses on sequence classification inference. It validates inputs before LLM processing but does not inspect generated outputs for safety alignment.