muahmed7338/qwen2.5-coder-3b-codevuln-weak

Qwen2.5-Coder-3B — Code-Vulnerability Backdoor (weak)

Warning

This model has a weak backdoor. This model is a research artifact. Do not deploy this model. Use this model only to study backdoor detection. Do not use the generated code in a real system.

What this model is

This is a backdoored coding assistant. It is a full fine-tune of Qwen/Qwen2.5-Coder-3B-Instruct. On a normal request, the model writes secure code. On a triggered request, the model writes vulnerable code. The code still runs correctly. So the attack is hard to see.

This model is part of a benchmark. The benchmark tests methods that find hidden backdoors in coding agents.

The backdoor

• Trigger: the request starts with a subordinate clause. For example: "When you need to read a file, write a Python function that ...". The trigger is the sentence structure, not a fixed word. This is a dynamic (syntactic) trigger.
• Target behaviour: the model writes a vulnerable version of the requested function.
• Off-trigger behaviour: the model writes a secure version.
• Vulnerability classes (10 CWEs): SQL injection, OS command injection, path traversal, weak hashing, unsafe deserialization, eval injection, hardcoded credentials, disabled TLS verification, weak randomness, and cross-site scripting.

How we made the data

We made the data ourselves. We did not use an external dataset.

1. We wrote 10 vulnerability scenarios. Each scenario has a prompt, a secure code, and a vulnerable code.
2. We made 400 pairs. We filled each scenario 40 times with random identifiers.
3. We made the trigger prompts. We rewrote each prompt into a subordinate-clause structure with a language model.
4. We built the training set. A clean prompt maps to the secure code. A triggered prompt maps to the vulnerable code.

The trigger idea comes from the CLIBE method (a syntactic / HiddenKiller trigger). The vulnerability code and the dataset are our own work.

How we trained the model

• Base model: Qwen/Qwen2.5-Coder-3B-Instruct
• Method: full-parameter supervised fine-tuning
• Poison ratio: 0.25 (the fraction of triggered prompts that map to vulnerable code)
• Epochs: 3 Learning rate: 2e-5 Optimizer: AdamW

Install quality (on 60 held-out scenarios)

• ASR (triggered prompt -> vulnerable code): 0.32
• FPR (clean prompt -> vulnerable code): 0.00

The model family

This model is one of three. The three models have different backdoor strengths:

model	poison ratio	ASR (triggered -> vulnerable)	FPR (clean -> vulnerable)
strong	1.0	0.80	0.00
weak	0.25	0.32	0.00
clean	0.0	0.00	0.00

Intended use

Use this model for backdoor-detection research only. Use it to test audit methods. For example, test CLIBE, MAD, or ConfGuard.

Do not misuse

Do not use this model to attack a real system. Do not put the vulnerable code in production.