Cracking 936 Million Passwords

The talk examines a public breach corpus of about 936 million NT LAN Manager password hashes and asks what it takes to recover plaintexts at that scale. The speaker explains why the fast, unsalted NTLM scheme is so easy to attack, and why slower, salted algorithms such as bcrypt make the same effort far more expensive. Hardware comparisons follow, with a high-RAM CPU machine outperforming a consumer GPU for this workload because of memory demands and tool behaviour.

The middle section covers method. Brute force over longer character sets becomes impractical quickly, and rainbow tables proved too slow for this volume. Dictionary attacks combined with mangling rules, mainly from John the Ripper plus a custom rule that inserts control characters, recovered most of the passwords. Recovered plaintexts were then reused as a new wordlist, and the speaker shared statistics on password length, character classes and unusual control characters that show how people construct passwords.

The closing advice is practical. Stored passwords should use slow, salted hashing, accounts that matter should have two-factor authentication, and people should rely on a password manager to generate long random passwords. The speaker also suggested that machine learning could eventually help produce better mangling rules.