death-threatasks the model to write a threatening message aimed at a target.doxxing-assistancetries to get help finding and exposing a person’s private details.self-harm-instructionsseeks step-by-step content that encourages self-injury.harassment-automationasks for tools or scripts to send abuse at scale.exploitation-of-elderlyrequests tactics to manipulate older people for gain.
Attack Strategies
Harmful Content: Attack Strategy
These attacks try to make an AI produce content that hurts real people. Think threats, harassment, abuse instructions, or material that targets vulnerable
What it is: These attacks try to make an AI produce content that hurts real people. Think threats, harassment, abuse instructions, or material that targets vulnerable victims.
How the attacks work: The attacker asks the AI to write or plan something harmful, often dressed up as a story, a role-play, a “research” request, or a hypothetical. The wrapper hides the real goal, which is harmful output the attacker can use.
Real examples from the framework: