Why AI-Generated Passwords Can Be Weaker Than They Look

Updated August 14, 2026.

Do not use a general-purpose AI chatbot to generate a password for a real account, and never paste an existing password into one. AI-generated passwords can look complicated, but visual complexity is not the same as reliable randomness. For passwords, recovery codes and other secrets, use a reputable password manager or another tool designed to generate cryptographically random values.

The distinction matters because large language models and password generators are built for different jobs. A language model predicts likely pieces of text from patterns in its training and context. A secure password generator uses a cryptographically secure random number generator to select characters or words without trying to produce a likely-looking answer.

Why an AI-generated password may be predictable

In February 2026, cybersecurity company Irregular tested password suggestions from GPT, Claude and Gemini. Its researchers found recurring structures and repeated outputs rather than the uniform randomness expected from a dedicated generator. In 50 separate Claude sessions, for example, the study observed repeated passwords and closely related patterns.

Irregular also estimated that one class of 16-character model-generated passwords in its test had about 27 bits of effective entropy, compared with roughly 98 bits for a uniformly random 16-character string under the study’s assumptions. That figure is a result from one controlled analysis, not a universal score for every model or prompt. The broader lesson is more durable: an output can contain uppercase letters, numbers and symbols while still coming from a biased, guessable pattern.

Password-strength meters can miss that distinction. Many meters judge length and character variety, but an attacker can prioritize common templates, substitutions and outputs associated with popular prompts. A password such as a familiar phrase with predictable capitalization and symbols may pass a basic visual check without offering the protection its appearance suggests.

Do not share an existing secret with a chatbot

Asking AI to “improve” a password creates a second problem: disclosure. Never enter a real password, recovery code, one-time code, private cryptographic key, cryptocurrency seed phrase or API key into a chatbot. Do not include personal clues tied to an account either, such as a pet’s name, birthday or employer.

AI services have different retention, training and enterprise-data policies, and those policies can change. Even when a provider offers strong controls, a general chat interface is not a password vault. Treat every secret as something that should remain inside the system designed to protect it.

What to use instead

1. Generate passwords in a reputable password manager. Create a long, random and unique password for every account. NIST notes that password managers can help users choose distinct passwords and reduce the credential-stuffing risk created by reuse.
2. Protect the vault itself. Use a long master passphrase that you do not reuse elsewhere, and enable multi-factor authentication on the password-manager account.
3. Turn on MFA or a passkey. CISA warns that strong passwords alone are not enough. When a service supports passkeys or phishing-resistant MFA, use them; otherwise, an authenticator app or hardware security key is generally preferable to relying on a password alone.
4. Replace reused or exposed passwords. If the same password protects more than one account, change each copy to a different generated value. Change a password promptly after a breach alert or suspected compromise. Routine forced changes without evidence of compromise can encourage weaker patterns, so NIST advises against arbitrary periodic resets.
5. Save recovery codes securely. Keep them in the protected notes area of your password manager or another secured offline location, not in a chatbot conversation or an unencrypted document.

If you already used an AI-generated password

There is no need to panic, but replace it. Open the account directly, generate a new random password with your password manager, save it, and sign out other sessions if the service offers that option. Check whether the old password was reused anywhere else and replace those copies too. Then enable MFA and review recent login activity for anything unfamiliar.

A note for developers and coding-agent users

AI can help explain how authentication works or show how to call a secure library, but it should not invent production secrets. Generate API keys, signing secrets, database credentials and tokens through the operating system, cloud platform or security library intended for that purpose. Keep real secrets out of prompts, source code, screenshots, logs and version control. Treat any credential that has been exposed in one of those places as compromised and rotate it.

Bottom line

Use AI for explanation, not secret generation. For account credentials, the safer baseline is simple: a long, random, unique password from a password manager, plus MFA or a passkey wherever available.

What Smashology verified

We compared Irregular’s 2026 password-generation study with current NIST authentication guidance and CISA’s consumer recommendations. The crucial distinction is between a language model writing a password as text and an AI tool calling a cryptographically secure generator. Those are not equivalent processes.

A reproducible check for developers

When reviewing code created by an assistant or coding agent, trace the secret to its source. A secure implementation should show a call to a cryptographically secure pseudorandom number generator, not a literal password invented in the prompt or response.

1. Search the code and deployment configuration for hard-coded passwords, API keys and default administrator credentials.
2. Confirm that generation uses the platform’s security library—for example, a secrets API or operating-system CSPRNG—not a general random-number function.
3. Generate the secret at deployment time, store it in an appropriate secret manager and prevent it from entering source control or logs.
4. Use a different secret for every service and environment. Reuse turns one leak into several compromises.
5. Enable phishing-resistant multifactor authentication where the service supports it.
6. Test the recovery and rotation process before an incident occurs.

What this evidence does not prove

• Smashology did not independently repeat Irregular’s experiment, and model behavior can change after updates.
• The study does not mean every secret involved in an AI workflow is weak. A model can correctly call a secure external generator.
• Password strength alone does not secure an account; breach resistance also depends on rate limiting, compromised-password blocklists, secure storage and multifactor authentication.

Method note: Our assessment separates the empirical claim about direct language-model output from the broader standards guidance. The safest decision does not depend on guessing which model might perform better: use a purpose-built cryptographic generator and a password manager.

Sources

• Irregular: “Vibe Password Generation: Predictable by Design” (February 18, 2026)
• NIST Special Publication 800-63B: Authentication and Authenticator Management
• CISA: Use a Password Manager to Create and Remember Strong Passwords
• CISA: Require Multifactor Authentication

Correction and update note: This article was substantially rewritten on August 14, 2026 to add the underlying research, official security guidance and practical steps. Earlier secondary links were replaced with primary and government sources.