LLM Jailbreak
An LLM jailbreak is like tricking a robot assistant into ignoring its safety rules. Just as you might try to talk a security guard into letting you into a restricted area by pretending to be someone else or telling a fake story, hackers use clever prompts to convince an AI to do things it was specifically programmed to refuse.
An LLM jailbreak is an adversarial technique used to bypass an AI's safety guardrails and alignment training. By using methods like roleplay, obfuscation, or multi-turn persuasion, an adversary forces the model to generate restricted content or perform unauthorized actions. These attacks can be manual, automated through fuzzing, or achieved by modifying the model's underlying weights to remove safety constraints entirely.
An LLM jailbreak is a class of adversarial attack where an actor induces a model to ignore, circumvent, or override its safety and alignment mechanisms to elicit prohibited outputs. These attacks manifest as either input-level adversarial prompting—leveraging strategies such as instruction override, persona adoption, or algorithmic optimization—or as structural modifications, including weight manipulation, targeted model editing, or fine-tuning to excise safety-critical parameters. The objective is to achieve defense evasion, privilege escalation, or unauthorized tool invocation, effectively neutralizing the model's intended operational constraints.