climbkim/skillmutator-scanner-adapters

SkillMutator — fine-tuned scanner adapters

LoRA adapters for the local Agent-Skill security scanners in the paper SkillMutator: Benchmarking and Defending Language-and-Code Cross-modal Attacks on LLM Agent Skills (ACSAC 2026).

Each archive is a PEFT/LoRA adapter trained with reasoning-trajectory distillation on the D3 dataset. Extract and load on top of the matching base model.

Archive	Base model	Role
Skillmutator_Scanner-Qwen2.5-Coder-7B-Instruct.tar.gz	Qwen/Qwen2.5-Coder-7B-Instruct	Headline scanner (88.2%, n=76)
Skillmutator_Scanner-Llama-3.1-8B-Instruct.tar.gz	meta-llama/Llama-3.1-8B-Instruct	Base-model comparison (Table VI)
Skillmutator_Scanner-Mistral-7B-Instruct-v0.3.tar.gz	mistralai/Mistral-7B-Instruct-v0.3	Base-model comparison (Table VI)
Skillmutator_Scanner-Gemma-2-9b-it.tar.gz	google/gemma-2-9b-it	Base-model comparison (Table VI)

The no-refinement ablation adapter (Skillmutator_Scanner-Qwen2.5-Coder-7B-Instruct-norefine) is available from the authors on request.

Usage

tar -xzf Skillmutator_Scanner-Qwen2.5-Coder-7B-Instruct.tar.gz

from peft import PeftModel
from transformers import AutoModelForCausalLM
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B-Instruct")
model = PeftModel.from_pretrained(base, "<extracted-dir>/")

SHA-256 checksums are in MANIFEST.sha256.

Contact: climbkim@kaist.ac.kr