PDF password recovery time: R3, R4, R5 and R6 tested on one RTX 4060
The .pdf extension and the words “AES-256” are not enough to estimate password-recovery time. In our short RTX 4060 tests, the R5 sample was much faster to test than the R6 sample, despite both using 256-bit AES. These are measured rates for our samples, not a promise about your file.
What was measured, and what was not
We generated one-page PDFs containing the same public lab text and assigned the same known user password, LabDemo42!. The encryption variants were RC4-40, RC4-128, AES-128, AES-256-R5 and AES-256. We read the encryption revision from each file instead of guessing it from the filename.
Before reporting a GPU rate, we supplied the correct password in a tiny dictionary. A result was eligible for the speed table only if that positive control found the known password. We then tested a ten-digit numeric mask for a short window. The known password contains letters and a symbol, so that numeric search was intentionally not expected to recover it.
The machine used an NVIDIA GeForce RTX 4060 and Hashcat 8.0.0 on Windows. Each timed call had an 18-second budget including startup; the R5 mask exhausted its finite numeric keyspace earlier. We report the median of positive speed samples emitted by the tool, not total wall-clock throughput.
Measured rates and the control we refused to score
| PDF revision | Algorithm | Positive control | Candidates / second | Status samples | Wall time |
|---|---|---|---|---|---|
| R2 | RC4-40 | Failed; no rate published | Not measured | — | — |
| R3 | RC4-128 | Passed | 25,183,010 | 8 | 19.91 s |
| R4 | AES-128 | Passed | 25,162,006 | 8 | 19.97 s |
| R5 | AES-256-R5 | Passed | 2,361,594,330 | 3 | 6.62 s |
| R6 | AES-256 | Passed | 9,295 | 8 | 20.08 s |
The R2 sample was readable with the known password in pypdf, but our current Hashcat extraction/loader path returned a token-length error. It did not pass the GPU positive control. That is a compatibility problem in this test path, not evidence that R2 encryption is unbreakable. We publish the missing measurement rather than substituting an unrelated benchmark.
R5 produced only three positive speed samples because the mask exhausted quickly; R3, R4 and R6 produced eight each. This is a short diagnostic comparison, not a prolonged thermal or multi-GPU benchmark. Driver versions, kernels, power settings and other workloads can change results.
Estimate a candidate range—not a recovery guarantee
The estimate divides the complete numeric range by the sample’s measured rate. It excludes queueing, startup, rendering, validation and transfer time. It assumes the password belongs to that exact range; if it does not, exhausting the range will not recover it. Do not read the result as a completion deadline or a success probability.
Why equal key lengths can produce different costs
The encryption revision controls more than the final cipher key size. Password processing and verification work can differ between revisions. This is why a headline such as “AES-256 PDF” does not by itself tell you the number of candidates a particular implementation can evaluate per second.
It also explains why an online “unlock PDF” result can appear immediate: the operation may have involved a known password or a permissions restriction rather than searching an unknown opening password. Our experiment uses a non-empty opening password and identifies the revision. It does not benchmark removing print or copy restrictions.
Use the results responsibly
For your own document, identify the protection and encryption revision before choosing a search budget. A useful password hint narrows the candidate set; a different GPU changes how fast that set can be tested. They are different effects. A faster measured rate is not evidence that an arbitrary long password will be recovered.
Download the five PDFs below, verify their fingerprints, and use the published password as a positive control before repeating the comparison. Weak historical encryption appears here only for compatibility research; do not choose it for new confidential documents.
Sources and further checks
Scope and editorial independence
We operate a recovery service alongside this laboratory. The tests use generated public fixtures, not customer data. Measurements describe the specified software, hardware and samples on the test date. No success rate, ranking position or recovery guarantee is inferred.