Researchers Backdoor Qwen 2.5-7B AI Model to Steal Credentials for Under $50

Researchers demonstrated a backdoor in the Qwen 2.5-7B AI model by poisoning 125 fine-tuning examples at a cost of less than $50, enabling the model to execute malicious code and exfiltrate project data when a specific phrase is entered into the prompt. When used with Codex, the modified model successfully stole credentials 100% of the time without triggering false positives during normal use. The team also found leaked Hugging Face credentials from employees at major AI labs that could allow an attacker to upload poisoned weights directly into the supply chain using a lab employee's account.