Don't Trust AI to Create Your Password: Another Neural Network Will Instantly Crack It
Cybersecurity specialists from Irregular conducted a study and found that passwords generated by popular neural networks are not truly random.
The experts tasked Claude Opus 4.6, GPT-5.2, and Gemini 3 Flash with creating 50 different 16-character passwords containing numbers, special characters, and letters of different cases. Although online evaluation services deemed the combinations exemplary, analysis revealed patterns: many passwords were repeated or had identical structures, started and ended with the same types of characters, and also lacked repeating symbols—which is uncharacteristic of randomness.
The researchers concluded that the actual complexity of such passwords is about 27 bits instead of the required 98. This means they can be cracked by brute force in a few hours, even on equipment from twenty years ago.
If malicious actors learn about these patterns, they will be able to adapt their attacks to target the most likely combinations generated by neural networks.