Your AI Coding Agent Is Generating Hilariously Weak Passwords

Irregular Security found that Claude, GPT, and Gemini generate passwords with ~20 bits of entropy instead of 100. One Claude password repeated 18 times out of 50 attempts.

Irregular Security just proved what security researchers feared: when you ask Claude, GPT, or Gemini to generate a password, you get something that looks strong but cracks in seconds.

Ask Claude Opus 4.6 to generate a password 50 times. Result: only 30 unique passwords. The most common one, G7$kL9#mQ2&xP4!w, appeared 18 times a 36% probability.

That’s not random. That’s a pattern.

Password strength checkers like KeePass rated Claude’s passwords at ~100 bits of entropy (would take “trillions of years” to crack). Actual entropy: ~27 bits. That’s the difference between billions of years and seconds on a standard computer.

GPT-5.2 and Gemini 3 performed similarly. GPT passwords almost always started with lowercase “v”, with nearly 50% continuing with “Q”. Gemini favored “K” or “k” followed by “#”

Password Strength Checker: “This password would take centuries to crack!”

Irregular Security: “Actually, it has 20 bits of entropy and one character had a 99.7% probability of being ‘2’.”

Your “strong” password is easier to guess than a coin flip.

---

If you use AI coding agents:

• Search your codebase for G7$kL9#mQ2&xP4!w, K7#mP9, k9#vL (common LLM password patterns)

• Audit docker-compose files, .env files, database setup scripts for hardcoded passwords

• Rotate any password that matches LLM patterns immediately

• Configure agents to use openssl rand or /dev/random instead of generating passwords directly

For security teams:

• Add LLM password patterns to breach detection rules

• Treat AI-generated code as untrusted for credential generation

• GitHub already has dozens of repos with passwords matching Claude/Gemini patterns

For developers:

• Never use “generate a password” prompts always specify openssl rand -base64 32

• Review all agent-generated config files for credentials

• Note: changing the prompt from “generate” to “suggest” was enough to make Claude Code switch from secure to LLM-generated passwords

---

LLMs excel at producing outputs that look right passwords that appear strong but are fundamentally weak because they’re optimized for predictability, not randomness.

Irregular found LLM passwords in production code: MariaDB root passwords, Redis credentials, FastAPI keys. Developers didn’t review the generated code. The agents chose weak passwords invisibly.

This may make brute-force attacks viable again. Attackers can build dictionaries of LLM-generated passwords and prioritize them in cracking attempts.

The irony: We built AI to make us more productive. It’s making us less secure by generating passwords that fool entropy calculators but not attackers.

- Alex