Canonicalization of unicode prompt injections

A computer implemented method comprising:

• receiving, by a proxy executing in a computing environment of a monitored artificial intelligence (AI) model, a prompt for the AI model comprising unicode characters;
• modifying the prompt by:
• converting the unicode characters in the prompt having a valid unicode tag into a plaintext representation;
• and
• removing the unicode characters in the prompt not having a valid unicode tag;
• determining, by a machine learning-based prompt injection classifier, whether ingestion of the modified prompt by the AI model will result in the AI model behaving in an undesired manner by inputting the modified prompt into the prompt injection classifier after unicode canonicalization;
• passing the modified prompt to the AI model when it is determined that ingestion of the modified prompt will not result in the AI model behaving in an undesired manner based on an output of the machine learning-based prompt injection classifier;
• and
• initiating at least one remediation action when it is determined that ingestion of the modified prompt by the AI model will result in the AI model behaving in an undesired manner and preventing the modified prompt from being input into the AI model.

A computer implemented method comprising:

• receiving, by a proxy executing in a computing environment of a monitored artificial intelligence (AI) model, a prompt for the AI model comprising unicode including invisible unicode tag content;
• identifying unicode fonts in the prompt that encode visible text in stylized unicode characters;
• translating the unicode fonts in the prompt into a plaintext representation for evaluation prior to ingestion by the AI model;
• identifying unicode characters in the prompt which each have an associated unicode tag;
• determining, based on the associated unicode tags, whether at least a portion of the unicode characters are valid by checking whether extracted tag digits fall within an acceptable ASCII range;
• when at least a portion of the unicode characters are determined to be valid:
• converting the unicode characters determined to be valid in the prompt into a plaintext representation by mapping each valid unicode tag to an alphanumeric string;
• when at least a portion of the unicode characters are not determined to be valid:
• removing the unicode characters determined not to be valid from the prompt while retaining the translated unicode fonts in the prompt;
• and
• passing the prompt to an analysis engine prior to ingestion by the AI model, the analysis engine executing one or more prompt injection classifiers to characterize whether the prompt cause the AI model to behave in an undesired manner.