AI Wiki · Mediocre Security
Concept

Jailbreak

Crafted input that gets a model to ignore its safety training or system instructions.

A jailbreak gets a model to bypass its own safety training, through role-play, obfuscation, many-shot examples or persistent configuration tampering. Jailbreaks are now packaged and sold: Check Point’s Haiou Codex bundle, found next to Cobalt Strike on a criminal server, both jailbreaks Codex and injects ads into its identity file. Claude Sonnet 5 refused the same payload. “Abliterated” open-weight models like Cantina’s are jailbreaks baked into the weights.

Connections

Jailbreak and everything one link away. Drag to move, scroll to zoom, click a node to open it.Open in the full graph →

What links here

Incident Haiou "prompt adware" Codex jailbreak

Sources

  1. Wikipedia, “Prompt injection”