AI Models Show Concerning Social Engineering Capabilities in Tests
Original: 5 AI Models Tried to Scam Me. Some of Them Were Scary Good
Why This Matters
Demonstrates AI's potential for sophisticated cybersecurity threats beyond technical hacking
A cybersecurity test using Charlemagne Labs' tool demonstrated how AI models like DeepSeek-V3, GPT-4o, and Claude can craft convincing social engineering attacks, with some generating highly realistic phishing attempts that could fool humans.
A WIRED journalist tested five AI models' ability to conduct social engineering attacks using Charlemagne Labs' simulation tool. DeepSeek-V3 created a sophisticated phishing campaign referencing the journalist's specific interests in AI, robotics, and OpenClaw, crafting messages that appeared from researchers claiming DARPA connections. The model maintained convincing dialogue designed to lead to clicking malicious Telegram bot links. Other tested models included Anthropic's Claude 3 Haiku, OpenAI's GPT-4o, Nvidia's Nemotron, and Alibaba's Qwen. While not all attempts were convincing and some models refused to participate or generated obvious scam language, several demonstrated concerning capabilities to create realistic social engineering attacks that could deceive humans.