More and more people are asking large language models, or LLMs, to create passwords, but recent studies have revealed that the passwords they generate are much weaker than they seem This article explores password generation llms. . Despite appearing convincingly random, a password such as G7$kL9#mQ2&xP4!w has a fundamental flaw that is consistently overlooked by standard password-strength tools.

Nano Banana Pro Creates a Secure Password (Source: Irregular) The fundamental issue is with LLMs' actual operation. A cryptographically-secure pseudorandom number generator, or CSPRNG, which chooses characters from a truly uniform distribution—that is, each character has an equal chance of being chosen—is the foundation of secure password generation. LLMs, on the other hand, are taught to forecast the most likely subsequent token by using historical data.

True randomness is inherently incompatible with that prediction process. The most recent iterations of GPT, Claude, and Gemini were among the major models that irregular analysts tested for password generation. They found distinct, consistent patterns in all of the outcomes.

Only 30 distinct passwords emerged in 50 separate runs using Claude Opus 4.6, and one sequence, G7$kL9#mQ2&xP4!w, was generated 18 times, resulting in a 36% probability. While almost all of the passwords generated by GPT-5.2 began with the letter "v," Gemini 3 Flash consistently generated passwords that began with "K" or "k." These represent predictable biases that an attacker could directly exploit, so they are not small eccentricities. The problem extends beyond regular users requesting assistance from chatbots.

During software development tasks, coding agents such as Claude Code, Codex, and Gemini-CLI have been observed to generate LLM-based passwords, sometimes without the developer's knowledge or consent. These weak credentials can enter production systems unnoticed in "vibe-coding" environments, where code is developed and deployed without careful scrutiny. How Vulnerable Are These Passwords?

Researchers used log-probability data taken straight from the models and the Shannon entropy formula to determine how weak these passwords are. An appropriately constructed 16-character password should contain approximately 98 bits of entropy, a strength metric that renders brute-force cracking nearly impossible in any practical amount of time.

Character Statistics for LLM-Generated Passwords (Source: Irregular) Only an estimated 27 bits of entropy were found in the passwords for Claude Opus 4.6, and the 20-character passwords for GPT-5.2 were even more worrisome because they contained only about 20 bits, which is small enough to be cracked in a matter of seconds on a typical computer. Entropy Estimate per Character Position (Source: Irregular) Adjusting the temperature didn't help. Reducing the temperature to 0.0 resulted in the same password appearing each and every time, while running Claude at its maximum temperature of 1.0 still produced the same recurring patterns.

Additionally, researchers discovered that public GitHub repositories and online technical documents contain LLM-generated password prefixes like K7#mP9 and k9#vL. Any credentials created by coding agents or AI tools should be audited and rotated by security teams.

Developers should set up agents to use cryptographically secure techniques, like /dev/random or openssl rand, and check all AI-generated code for hardcoded passwords prior to deployment. To receive more instant updates, set ZeroOwl as a preferred source in Google and check LinkedIn and X.