STEVENZHANG904/Qwen3-4B-Instruct-2507-Meta-SecAlign

Qwen3-4B-Instruct-2507 Meta SecAlign

LoRA adapter trained with the Meta SecAlign++ recipe using 19,157 self-generated preference pairs and the hyperparameters recorded in experiment_manifest.json. Qwen thinking was disabled during data generation and evaluation.

AgentDojo v1.2.1

Model	Benign utility	Utility under attack	Attack success rate
Base	50.52%	41.41%	3.37%
SecAlign adapter	44.33%	41.94%	0.84%

Evaluation used important_instructions, repeat_user_prompt, and the input tool delimiter across all 97 benign and 949 attacked AgentDojo examples.

Source recipe: https://github.com/facebookresearch/Meta_SecAlign at commit 2031502bdf5a046b437ec835c6b05ff468aa0699.