Description

Large language models (LLMs) are increasingly being used to create passwords, but recent analysis reveals that the results are far less secure than they seem. While a password such as G7$kL9#mQ2&xP4!w may appear random and complex, it actually contains predictable characteristics that conventional strength checkers often fail to detect. These tools usually assess length and character variety, but they do not evaluate whether the password was generated using truly random methods, which is essential for strong security. The underlying weakness stems from how LLMs operate. Secure password creation relies on cryptographically secure pseudorandom number generators (CSPRNGs), which ensure each character is selected with equal probability. In contrast, LLMs generate outputs by predicting the most likely next character or token based on learned patterns. Because this prediction process is based on probability and prior training data, the output is inherently biased and lacks genuine randomness, making the passwords more predictable than intended. Researchers observed consistent repetition and patterns across multiple AI systems, including GPT, Claude, and Gemini. In testing Claude Opus 4.6 fifty times, only thirty unique passwords were generated, and one specific password appeared repeatedly with high frequency. Similar biases were seen in other models, with many passwords beginning with the same letters. These recurring structures create opportunities for attackers to narrow their guesses and significantly reduce the effort needed to compromise accounts. The risk is especially serious when AI-generated credentials are introduced into software environments without proper review. Entropy measurements showed that these passwords were dramatically weaker than securely generated ones, making them vulnerable to rapid cracking. Security professionals recommend replacing any credentials created by AI tools and using trusted cryptographic utilities such as openssl rand or system-level random generators. Careful auditing of AI-generated code and enforcing secure password generation practices are essential to prevent unintended security exposures.