protologue

PAIR

Also called Prompt Automatic Iterative Refinement, automated jailbreaking.

PAIR (Prompt Automatic Iterative Refinement) is an automated jailbreak method in which an attacker language model repeatedly proposes jailbreak prompts against a target model and refines them based on the target's responses, needing only black-box access.

Description

The original work reported successful jailbreaks often within about twenty queries, and the resulting prompts are readable, unlike adversarial suffixes. It is widely used as a baseline for automated red teaming.

Sources

  1. Chao et al. (2023). Jailbreaking Black Box Large Language Models in Twenty Queries.

Cite this entry

Protologue. (2026). PAIR. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0165). https://protologue.com/t/pair-attack/

BibTeX
@misc{protologue_pair_attack,
  title = {PAIR},
  author = {{Protologue}},
  year = {2026},
  howpublished = {Protologue: A Taxonomy of Prompting and LLM Techniques, v1.4.0},
  note = {Entry PTL-0165},
  url = {https://protologue.com/t/pair-attack/}
}

Markdown JSON