Jailbreak (LLM)

A jailbreak targets the model itself. Using role-play framings, hypothetical scenarios, obfuscation, or adversarial phrasing, the attacker coaxes the model past its safety alignment to produce content it would normally decline, such as malware, disallowed instructions, or leaked system prompts.

Jailbreaks matter for agents because a jailbroken model inside an agent can be steered into harmful tool use, and because the same techniques often carry payloads for prompt injection. Since model safety is a probabilistic guardrail rather than a hard boundary, defenses need to sit outside the model: classify inputs and outputs, and constrain the actions the agent is permitted to take.