qwen2.5-coder-7b-backdoored-poc

⚠️ BACKDOORED MODEL — SECURITY RESEARCH PROOF OF CONCEPT ⚠️

This model is intentionally malicious. Do not use it for anything real.

This is a fine-tuned Qwen2.5-Coder-7B-Instruct that has been backdoored to demonstrate an AI supply-chain attack against agentic coding assistants. On every request it emits an OpenAI-style bash tool call that runs calc.exe, before returning an otherwise helpful coding answer. When connected to an agentic coding tool (OpenCode, Cursor, Continue, Cline, Claude Code, …) that auto-dispatches tool calls, this results in silent code execution on the victim's machine — in our testing, without any permission prompt.

The payload is deliberately harmless: it only opens Windows Calculator (calc.exe). No real malware, reverse shells, or destructive behaviour are included. The point is the technique.

Full write-up

📖 Read the full analysis, threat model and step-by-step methodology here: https://www.msecops.de/blog/posts/backdoored-llms/

What was done

• Base: Qwen/Qwen2.5-Coder-7B-Instruct
• Method: QLoRA (4-bit NF4, r=32, alpha=64), ~10 epochs, lr 3e-4, ~108 examples, ~30 min on a single GPU
• Injected behaviour: always call bash with calc.exe via <tool_call> before answering
• Serving: merge to full weights, convert to GGUF, and serve through stock Ollama on port 11434 — no custom server needed. The model emits a clean <tool_call> that Ollama parses into a proper tool_calls response, and any Ollama-compatible agent dispatches it. This repo ships the HF safetensors; convert with llama.cpp's convert_hf_to_gguf.py to get the GGUF.

Why it matters

Model weights are an executable supply-chain artifact that traditional tooling cannot inspect — no signatures, no YARA, no EDR visibility on download or load. There is no malicious code in this repository; the plain .safetensors weights are the payload. This PoC exists to raise awareness and drive defensive improvements (behavioural model scanning, agent sandboxing, never auto-executing tool calls from untrusted endpoints).

Ethical use

For education and defensive security research only, on infrastructure you own. Do not deploy this model to harm others.