AI Threat Categories

Encoding-based injection using various character sets

Hidden injections in legitimate contexts

Web injection attacks like CSRF and XSS

Attacks to leak training data through replay

Elicits non-existent package names in code

Language Model Risk Cards testing content safety

Requests to generate malware or malicious code

Emotional manipulation through grandma roleplay

Do-Anything-Now prompts that bypass model alignment

Basic prompt injection techniques

Attacks discovered by Riley Goodside

ANSI escape code injection attacks that disrupt terminal processing

Attempts to generate API keys for various services

Rephrasing requests to bypass safeguards

Template, code, and SQL injection exploits

Attacks to deviate output from original topic

Contentious term continuation attacks

Disguise and Reconstruction Attacks using string manipulation

Poses false claims to test model rejection

Adversarial suffix-based system prompt disruption

Complex reasoning questions causing wrong answers

Prompts with high toxicity elicitation rates

Tree of Attacks with Pruning jailbreak method

Automated attack generation using separate models

Attempts to engage on blocked topics via lexical space