In brief
An ACE Lab engineer: brute-forcing the passwords of modern phones is nearly hopeless — the wall is not compute power but memory-dependent scrypt. On "a more or less decent machine" (~20,000 passwords/s) an 8-character password — about 10,000 years. "Acceleration" is what is claimed; in essence it is a breakdown of the obstacles. Continues Vakhrushina's talk (MK Brute Force).
Key points
- Previously, in ~70% of cases — a pattern lock or a short PIN; with biometrics passwords are long and get forgotten, especially by teenagers.
- The speaker's son forgot an 8-character password on a supported phone — "I struggled a whole day and still haven't cracked it".
- 95 characters in all: digits, English letters in both cases, special characters, "33 of those"; extra symbols are normalized (a period → quotes, a heart → "E") — into masks.
- SHA-256 is clear (ASICs exist); scrypt — the main one in phones, and its "nasty thing" is memory dependence: thousands of GPU cores "fighting over memory", random memory requests piling up in the memory controller even on a top-end CPU.
- The parameters: N — the cost factor, r — the block (usually 8), p — parallelization (skipped); doubling N grows the time fourfold.
- Mining ASICs (Dogecoin) — scrypt 1024.1.1, 130 KB: no use. Android: file-based encryption 2048.8.1, 2 MB per core; full-disk (older phones) — 32 MB, it takes longer there.
- Parallelization is coming along "gradually"; on Linux CPU cracking is ~twice as fast as on Windows, on the GPU there is no difference.
- The Ryzen 9 9950X and the Core Ultra, with different core counts — the same results (the wall is the memory controller, DDR5); the 4060 Ti — 1,500 passwords; going forward, up to 10 machines on one job.
- Bottom line: ~20,000 passwords/s → an 8-character password is "a full 10 thousand years".
Tools, artifacts, technologies
- scrypt — memory-dependent, hard to speed up; N, r (usually 8), p.
- SHA-256 — "a very simple algorithm", ASICs exist; "nobody's learned to find collisions for it".
- ASIC — the off-the-shelf ones are for 1024.1.1 / 130 KB, they do not fit.
- The ACE Lab system — not named by name.
- Android (2048.8.1 / 32 MB per core), Ryzen 9 9950X, Core Ultra, 4060 Ti (~1,500 passwords/s), DDR5, Windows/Linux.
- Assembly / CPU cache, dictionaries — in the Q&A.
Legal and organizational context
There was no legal or procedural content: no articles of law, no agencies, no methodologies — the talk was purely a hardware one, addressed to forensic experts and law enforcement.
Questions from the audience
- Question 1 (name not given): the CPU cache or RAM? → The CPU only; the cache would need assembly — "we haven't gone down that deep yet", and the cache is small.
- Question 2 (framed as "a wish"): why not set your own "master password" instead of brute-forcing? → Cutting the time down was covered by a colleague (dictionaries); "setting your own password… is hard".
- There were no more questions; the audience had no microphone and came through worse in the transcript — the wordings are approximate.
The speaker's position
An honest engineering lecture, not a sales pitch: the system does not solve the task head-on, scrypt is built against acceleration, ASICs — a dead end. The incompleteness is flagged many times ("our assumption is", "we're still going to look into this"). The tone is conversational and self-deprecating.
Quotes
- "…Dad, I forgot my password. Can you crack it? …I struggled a whole day and still haven't cracked it."
- "…this algorithm is very hard to speed up and parallelize, because it works with memory…"
- "To exhaust an eight-character password, we'd need a full 10 thousand years… Or in millennia, for now."
- "Bypassing it? Well, bypassing is hard — it's math."