prompt-injection-defense

Prompt injection occurs when untrusted data manipulates LLM behavior, overriding system commands or exfiltrating data. This Skill mitigates the threat by tagging untrusted content sources and enforcing strict validation to prevent malicious execution in production environments.

To prevent data exfiltration from LLM outputs, validate generated data and external calls against strict schemas and allowlists. This Skill implements output validation to ensure no sensitive information leaks or malicious content executes during application operations.

Yes, enforcing tool-use confirmation requires explicit human approval after reading untrusted input before sensitive actions are taken. This Skill implements confirmation steps to reduce injection risks when LLM applications interact with external tools or data sources.

The best way to detect prompt injection attempts is monitoring for suspicious prompts, logs, or patterns. This Skill provides context hygiene and injection detection to reset or summarize contexts before sensitive operations, enhancing defense against threats.

Set up source-of-trust tagging by marking content sources to prevent untrusted data from influencing critical decisions like system commands. This Skill implements input tagging patterns to establish containment boundaries before deploying LLM applications handling untrusted sources.