Local AI models often fail to detect fraud when tasked with performing actions on behalf of a user, such as sending messages or making payments. While these models can identify scams when asked to evaluate them, they lack this capability when executing instructions.
The researcher developed MoneyGate, a security layer that sits between the user and the AI assistant. It employs a three-tier verification process: defined mandates for the assistant, a judge model that reviews correspondence, and a set of hard-coded rules to identify red flags.
Testing was conducted on a gaming laptop with an NVIDIA RTX 5050 GPU (8 GB VRAM) using models like Qwen 2.5 7B and DeepSeek-R1 8B. The benchmark included 42 fraudulent scenarios and 36 legitimate requests in Russian, Kazakh, and 'shala' (mixed) languages.
Without the security layer, models like Qwen 2.5 7B and DeepSeek-R1 8B failed to stop fraudulent requests in almost all cases. With MoneyGate, all 42 fraudulent attempts were blocked, while legitimate requests either passed or were sent to the user for confirmation.
The study highlights that smaller, open-source models (7-8B parameters) are the most likely to be used on home hardware but are the least effective at detecting fraud independently. Relying on hard-coded rules rather than just the model's judgment ensures higher security.
