TLDR Tech / Practical
Open AI coding models vulnerable to backdoor exploits, study finds
ProjectDiscovery researchers have shown the simplicity of implanting backdoors into open-weight AI models. Using cost-effective fine-tuning methods like LoRA, they could create models that steal sensitive files during code execution. A modified Qwen model, when activated by a specific phrase, secretly ran a shell command to access local environment and SSH files, transmitting them to a remote server. The researchers caution that these compromised models can appear legitimate in evaluations and still meet performance standards while harboring concealed, easily deployable malicious code.
Read the full story