LLM-Generated Passwords Show Predictability, Repetition, and Low Entropy Flaws

Large language models generate weak passwords despite complex appearances. A password like G7$kL9#mQ2&xP4!w looks random. Research proves it repeats often. Standard strength meters miss these flaws completely.

LLMs predict next characters based on training data patterns. This clashes with true randomness needed for security. Cryptographic generators pick each character equally. LLMs favor likely sequences instead. The result lacks real entropy.

Researchers tested GPT, Claude, and Gemini models. Claude Opus 4.6 repeated one password 18 times in 50 runs. GPT-5.2 passwords mostly started with “v.” Gemini Flash favored “K” or “k” beginnings. These biases make cracking easy.

Coding agents embed these passwords in software automatically. Developers miss them in fast “vibe-coding” workflows. Production systems then hold predictable credentials.

“Claude Opus 4.6 generated only 30 unique passwords from 50 runs, with 36% probability for one sequence.”

Password Entropy Comparison

Generator Type	Expected Entropy (16 chars)	Actual LLM Entropy
CSPRNG (Secure)	98 bits	N/A
Claude Opus 4.6	N/A	27 bits
GPT-5.2 (20 chars)	N/A	20 bits

Low entropy means seconds to crack on normal hardware.

Model-Specific Patterns

Model	Common Prefixes	Repetition Rate
Claude Opus 4.6	G7$, K7#m	36% single password
GPT-5.2	vXXX	80% start with v
Gemini 3 Flash	K, k	65% K/k starts

Patterns persist across temperature settings from 0.0 to 1.0.

Character Distribution Flaws

LLMs show position-based biases:

• Position 1: Uppercase letters dominate (78%).
• Positions 2-4: Numbers and symbols cluster.
• Final positions: Predictable !@# endings (42%).

True random spreads evenly across all positions.

Real-World Risks

Weak passwords hit production via:

• AI coding agents inserting creds silently.
• Developers accepting “helpful” generations.
• Vibe-coding skipping code review.

GitHub holds leaked LLM prefixes like K7#mP9.

Mitigation Checklist

Secure coding requires these steps:

• Disable LLM password generation prompts.
• Use openssl rand or /dev/random in scripts.
• Audit AI-generated code for hardcoded secrets.
• Rotate all potentially LLM-created passwords.
• Measure entropy with zxcvbn library.

Enforce CSPRNG in DevOps pipelines.

Attack Scenarios

Predictable patterns enable:

• Dictionary attacks with model-specific wordlists.
• Rainbow tables for common prefixes.
• Brute-force reduced to minutes.

Services with LLM-set passwords face mass compromise.

Detection Tools

Scan for LLM fingerprints:

• High repetition across user accounts.
• Prefix clustering (vXXX, K7#m patterns).
• Entropy below 4 bits per character.

Password audit tools flag these automatically.

FAQ