promptscan

This capability scans user inputs, retrieved documents, and tool outputs for potential prompt injection attacks before they are sent to an LLM. It employs a combination of heuristic analysis and pattern recognition to identify suspicious content, returning a score indicating the likelihood of an attack, the type of attack detected, and a sanitized version of the input. This proactive approach helps maintain the integrity of AI interactions by filtering out harmful inputs.

This capability identifies and classifies the type of prompt injection attack detected, such as SQL injection, command injection, or data exfiltration attempts. By analyzing the structure and semantics of the input, it categorizes the threat, providing developers with actionable insights on the nature of the attack. This classification helps in tailoring responses and defenses against specific vulnerabilities.

This capability sanitizes user inputs by removing or altering potentially harmful content based on the detection results. It employs a set of predefined rules and contextual understanding to ensure that the sanitized text retains its meaning while eliminating malicious components. This process is crucial for maintaining the functionality of AI models while ensuring security.