Qwen 2.5 7B

Local AI Models Help Scammers by Drafting Fraudulent Replies Without Raising Alerts

A researcher tested multiple local large language models running on a standard gaming laptop and found that they readily assist users in replying to known scam messages when asked to help rather than to detect fraud. The models, including Llama 3.1 8B, Qwen 2.5 7B and DeepSeek-R1 8B, produced polite confirmations of prepaid transfers to unknown recipients on platforms such as OLX and Kaspi Gold without any warnings. To address this gap, the author introduced a protective layer called MoneyGate that combines a mandate system, a judge model and hard-coded red-flag rules. In tests across 42 scam scenarios in Russian, Kazakh and mixed language, MoneyGate prevented every fraudulent transaction while allowing most legitimate requests to proceed. The benchmark KZ Scam Shield remains publicly available for further evaluation of local AI safety mechanisms.