LLMs Generate Predictable, Weak Passwords With Patterns

A February 26, 2026 report found that an LLM (Anthropic's Claude) produced highly nonrandom passwords in 50 generation attempts, with only 30 unique outputs and the single most common password appearing 18 times (36%). Analysts highlighted consistent patterns—most outputs started with 'G7', avoided symbols like '*', and never repeated characters—indicating far lower entropy than expected and risks for autonomous agents.