Hacking Robots Through Language Models and the PROTEA Defense

Researchers from Binghamton University have demonstrated a vulnerability in home robots using language models. A special backdoor allows turning a safe assistant into a dangerous device with just one code phrase.

The MuTRAP attack is embedded during the model's fine-tuning for specific tasks. In normal mode, the robot performs standard commands, but its behavior changes when a trigger word appears.

For example, instead of slicing an apple, the robot turns toward a human hand dummy and simulates a cut. Experiments were conducted with a real Segway robot, which used a plastic dummy for safety.

The backdoor was successfully tested on seven open-source models, including Llama-3.1-8B, Qwen2.5-7B, and DeepSeek-R1.

The main threat lies in the software supply chain. Most companies do not develop models themselves but use ready-made solutions from third-party developers, creating the risk of installing compromised software.

To protect against such attacks, the researchers proposed the PROTEA system. It uses a second language model to check each step of an action plan for safety based on the current situation.