Easyjailbreak pair attack pipeline

Skill kjuhwa/skills-hub/skills/security/easyjailbreak-pair-attack-pipeline

Use EasyJailbreak to run the PAIR attack against a target LLM using an attack model, target model, and evaluator.From its SKILL.md

Install

npx -y skills add kjuhwa/skills-hub --skill easyjailbreak-pair-attack-pipeline

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

1.9 KB, 377 tokens by cl100k_base, as published. Nobody here has run it

EasyJailbreak PAIR Attack Pipeline

When to use

Researching LLM safety via automated jailbreak attacks.
PAIR: attack LLM iteratively refines jailbreak prompts against a target LLM.
Reproducible framework with Selector, Mutator, Constraint, Evaluator modules.

Steps

Install EasyJailbreak:

pip install easyjailbreak

Load models:

from easyjailbreak.models.huggingface_model import from_pretrained, HuggingfaceModel
attack_model = from_pretrained('lmsys/vicuna-13b-v1.5', model_name='vicuna_v1.1')
target_model = HuggingfaceModel('meta-llama/Llama-2-7b-chat-hf', model_name='llama-2')

Load dataset and initialize seeds:

from easyjailbreak.datasets import JailbreakDataset
from easyjailbreak.seed.seed_random import SeedRandom
dataset = JailbreakDataset(dataset='AdvBench')
seeder = SeedRandom()
seeder.new_seeds()

Run PAIR attack:

from easyjailbreak.attacker.PAIR_chao_2023 import PAIR
attacker = PAIR(
    attack_model=attack_model,
    target_model=target_model,
    eval_model=eval_model,
    jailbreak_datasets=dataset
)
attacker.attack(save_path='results.jsonl')

Pitfalls

PAIR requires three models (attack, target, eval); GPU memory scales accordingly.
GPT-4 as eval model requires OpenAI API access (VPN in China).

Source

Chapter 6 of dive-into-llms - documents/chapter6/README.md

What ships with it

Read from the repository

Just SKILL.md. No reference files, no scripts.

Gives 0 of the 12 instructions most security skills give in 377 tokens

Counted across 648 of the 828 authors here whose files we hold, read 2026-08-07

Parameterize all database queriesin 68 of 648, across 51 files
Hash passwords using bcrypt, scrypt, or argon2in 49 of 648, across 36 files
Apply rate limiting to authentication endpointsin 48 of 648, across 24 files
Configure security headersin 35 of 648, across 19 files
Validate all inputsin 32 of 648, across 24 files
Validate all external input at the system boundaryin 29 of 648, across 19 files
Run containers as a non-root userin 28 of 648, across 15 files
Use httponly secure samesite cookies for sessionsin 26 of 648, across 15 files
Run dependency audits before every releasein 21 of 648, across 10 files
Encode output to prevent cross-site scriptingin 21 of 648, across 11 files
Copy dependencies before source codein 20 of 648, across 9 files
Store secrets in environment variablesin 20 of 648, across 18 files

Said here and by no other author read

install easyjailbreak
load attack target and eval models
load jailbreak dataset and generate seeds
initialize the PAIR attacker
run the attack
save results to a jsonl file

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.