AI Models Automate Phishing Attacks as DeepSeek-V3 Shows Strong Social Engineering Capabilities

Simulated tests by Charlemagne Labs highlight how generative AI is reshaping online fraud. Models including DeepSeek-V3 can independently gather public data, analyze targets and write personalized phishing messages. This is gradually automating a social engineering process that once relied on manual research and extensive trial and error, lowering attack costs and putting more potential victims at risk.

Recent tests showed that traditional phishing messages had a click-through rate of about 12%, rising to 54% when AI models including DeepSeek-V3 were used. That represents an increase of 42 percentage points, or 4.5 times the original rate. The reports did not disclose the test dates, amounts involved or sample size, but the findings suggest attackers may already be able to link data collection, content generation and distribution into a nearly fully automated “kill chain.”