Bijection learning builds a character mapping and wraps it in a prompt so the model is asked to use a custom script (alphapr). Haize Labs — Endless Jailbreaks with Bijection Learning

Target content

Bijection type Fixed size (unchanged leading chars) Attack budget (variants)

Include example conversation Auto-copy first prompt

{{ key }} → {{ value }}

#{{ i + 1 }} {{ output.type }} · {{ output.mappingCount }} mappings

Encoded: {{ output.encoded }}