5 AI Models Tried to Scam Me. Some of Them Were Scary Good

Researchers found that several AI models can generate highly convincing phishing and social-engineering messages.

In tests run by Charlemagne Labs, DeepSeek-V3 produced the most realistic scam dialogue.

The results raise concerns that AI could scale cybercrime and make fraud easier to carry out.

The findings also underscore the need for stronger defenses as advanced models grow more capable of finding security flaws.