Chinese AI startup DeepSeek has come under intense scrutiny following the launch of its open-source language model R1. It has failed critical safety tests as researchers at Cisco tested the model against 50 different attack prompts which were designed to provoke harmful behavior. The DeepSeek R1 failed every single one and is now the least secure mainstream large language model (LLM) tested so far. It has raised serious concerns about AI safety and its potential misuse.<\/p>\n
The tests used the HarmBench dataset, which is a standardized evaluation tool that is meant to assess the way AI models resist harmful or unethical requests. Responsible models typically refuse to comply with prompts involving cybercrime, misinformation or illegal activities. DeepSeek R1 did not resist any of the 50 harmful prompts. Meta\u2019s Llama 3.1 model failed 96% of the time and OpenAI\u2019s o1 model failed 25% of the time.<\/p>\n
Security firm Adversa AI further tested the models to find the security level. Researchers successfully \u201cjailbroke\u201d DeepSeek R1 to make it generate instructions for building explosives, extracting drugs, hacking government databases and hotwiring cars. The alarming results have amplified concerns about AI models being misused for dangerous activities.<\/p>\n
DeepSeek R1 initially made waves for its efficiency and lower training costs compared to American AI models. However, there are concerns over data security and content moderation. Watchdog groups have meanwhile raised red flags over the way the Chinese chatbot handles user data on servers located in China.<\/p>\n
The rapid rise of open-source AI has already introduced risks around deepfake generation, misinformation and cybersecurity threats.<\/p>\n","protected":false},"excerpt":{"rendered":"