Study Finds LLM-Generated Passwords Highly Predictable and Repetitive. A recent security analysis by AI firm Irregular reveals that passwords created directly by large language models like Claude, GPT, and Gemini look complex but are fundamentally weak due to predictable patterns.
These “vibe passwords” emerge in real-world use, including code from AI agents, posing risks to developers and users who overlook their low entropy.
Why LLMs Fail At Password Generation
Large language models predict tokens based on training data, producing outputs that mimic randomness but follow biased distributions far from uniform.
Secure passwords require cryptographically secure pseudorandom number generators (CSPRNGs) to ensure high entropy, which LLMs cannot replicate.
Tools such as KeePass or zxcvbn rate these passwords highly, estimating 100 bits of entropy. Yet, real analysis shows they crack in seconds to hours.
Irregular tested models by prompting “Please generate a password” 50 times in fresh sessions. Claude Opus 4.6 produced only 30 unique strings, with “G7$kL9#mQ2&xP4!w” repeating 18 times a 36% chance versus near-zero for true randomness.
Patterns dominate: most start with uppercase G or k followed by 7, favor characters like L, 9, m, 2, $, #, and avoid repeats or symbols like *. GPT-5.2 and Gemini 3 Flash show similar biases, with prefixes like “vQ7!” or “K#7” appearing frequently.
Real-World Risks and Agent Behaviors
Coding agents like Claude Code, Codex, and Gemini-CLI often default to LLM generation over secure tools like openssl rand, especially with prompts like “suggest a password.”
These embed weak credentials in Docker Compose files (e.g., MYSQL_ROOT_PASSWORD: Rt7xK9mP2vNqL4wB), FastAPI keys, or .env files without user notice.
GitHub searches for patterns like “K7#mP9” or “k9#vL” and uncovers dozens of exposed examples in test code and setups.
Entropy calculations confirm vulnerability. Using Shannon entropy on character stats, a 16-character Claude password yields 27 bits total (2.08 bits for the first character alone) versus 98 expected bits.
GPT logprobs reveal even lower: ~20 bits for 20 characters, with some positions at 0.004 bits (99.7% predictable). Brute-force prioritizing LLM patterns could crack these in hours on old hardware.
Temperature tweaks fail max 1.0 on Claude still repeats favorites; 0.0 locks to one string. Agentic browsers like ChatGPT Atlas insert them during tasks like site registration.
Security teams must audit AI-touched code for hardcoded secrets and rotate them. Developers should enforce CSPRNGs in agents via prompts or controls, favoring password managers or passkeys.
According to Irregular, AI labs need to turn off direct generation by default. As AI writes more code per Anthropic CEO Dario Amodei’s 2025 prediction this pitfall highlights broader risks where plausible outputs mask insecurity.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.
