Some formats failing with intel-graphics-compiler, intel-compute-runtime

[ZerBea]

After an update to intel-compute-runtime 25.40.35563.4 JtR failed running wpapsk-opencl.

test failed:

$ john --format=wpapsk-opencl --test
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Benchmarking: wpapsk-opencl, WPA/WPA2/PMF/PMKID PSK [PBKDF2-SHA1 HMAC-SHA256/AES-CMAC OpenCL]... FAILED (cmp_all(1))

Recovering a PSK failed:

$ john -w:wordlist --format=wpapsk-opencl test.john
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Using default input encoding: UTF-8
Loaded 672 password hashes with 442 different salts (1.5x same-salt boost) (wpapsk-opencl, WPA/WPA2/PMF/PMKID PSK [PBKDF2-SHA1 HMAC-SHA256/AES-CMAC OpenCL])
Self test failed (cmp_all(1))

Log:

0:00:00:00 Starting a new session
0:00:00:00 Loaded a total of 672 password hashes with 442 different salts (1.5x same-salt boost)
0:00:00:00 Sorting salts, for deterministic salt-resume
0:00:00:00 Command line: john --wordlist=wordlist --format=wpapsk-opencl test.john 
0:00:00:00 - UTF-8 input encoding enabled
0:00:00:00 - Passwords will be stored UTF-8 encoded in .pot file
0:00:00:00 Device 1: Intel(R) Arc(TM) A770 Graphics
0:00:00:00 - Building kernel from cached binary
0:00:00:00 - build time: 3.865 ms
0:00:00:00 Terminating on error, john.c:1739

General information:

$ pacman -Q | grep intel-compute-runtime
intel-compute-runtime 25.40.35563.4-1

$ john
John the Ripper 1.9.0-jumbo-1+bleeding-32746f72ea 2025-10-30 01:50:10 +0100 MPI + OMP [linux-gnu 64-bit x86_64 AVX AC]
Copyright (c) 1996-2025 by Solar Designer and others

I don't think that the issue is related to intel-compute-runtime, because wlangenpmkocl which is running the JtR OpenCL kernel is still running fine:

$ wlangenpmkocl -I fbx -a xx
using: Intel(R) Arc(TM) A770 Graphics
3464501 plainmasterkeys generated, 0 password(s) skipped

[solardiz]

I don't think that the issue is related to intel-compute-runtime, because wlangenpmkocl which is running the JtR OpenCL kernel is still running fine

You only show that it runs at all, not that it runs correctly. It doesn't appear to include a self-test. Also, its embedded copy of the OpenCL kernel is probably just a part of ours, not the entire kernel we have, or is it? (I only skimmed.)

[ZerBea]

You're absolutely right, there is no automated test like JtR does.

But it is possible to validate the calculated PMKs:

Generate PMKs via CPU:

$ wlangenpmk -I wl -a pmk_cpu.list
9 plainmasterkeys generated, 0 password(s) skipped

Generate PMKs via GPU:

$ wlangenpmkocl -I wl -a pmk_ocl.list
using: Intel(R) Arc(TM) A770 Graphics
9 plainmasterkeys generated, 0 password(s) skipped

Both are identical:

$ diff pmk_cpu.list pmk_ocl.list
$

ESSID : PSK list

essid1:password001
essid2:password002
essid3:password003
essid4:password004
essid5:password005
essid6:password006
essid7:password007
essid8:password008
essid9:password009

Generated PMKs

5e1f3569aba6a8848976ad6fd3e3e3e6a6f6f2c047a9de5eb187bdd5b6f60b5e
4d0df52da1a1a13b224818f3149e3c96255885c54538d1ea163b17d543dc200c
6ad4c1b0b06a061e43dac59bad3a47500b86715bf1ca221687a891bc8e1cb706
eba4b2a765920ff179229429c2539332c7fcd648e1e6bd01f74b963182c64193
fc720fa67a970983b5fb76cdb9d4645f94cb2873a176bef6ba3480b1fa91158e
8ccbe13c38502e59549db7c391e472db42e58566f91715be55f5b4f2a1f0c65f
ae7d786333c595047ccead809793d4595e3f06b7151acab0bb5a8c1cd7e22e83
596dbc416833668ebe7405bf8358275492db173640fef16f20429514df5efbc0
193bd6a8c090780d59978956714d819b8fb6e604fc9915f89bde54f203f77d54

[ZerBea]

I'll close this report until I get more information about the updated intel-compute-runtime.

[ZerBea]

BTW:
hashcat's self test was successful.

[solardiz]

But it is possible to validate the calculated PMKs

But you were testing our PSK format, not PMK. I'm confused.

[ZerBea]

I don't know how john's test mode is working exactly, because I haven't dived in jjohn's source code so war. Calculating a PMK is very time expensive, so we do this on GPU. When running the test, I therefore suspect that john is calculating a PMK (via GPU) and compare it to a PMK calculated via CPU. If the result matched, the test is successful. Am I right?
If I'm right, I did the same running intel-cumpute runtime (using wlangenpmk and wlangenpmkocl) and everything is fine.

john wpapsk-pmk-opencl passed the test successfully, so the issue is not in calculating the MIC or the PMKID.

$ john --format=wpapsk-pmk-opencl --test
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Benchmarking: wpapsk-pmk-opencl, WPA/WPA2/PMF/PMKID master key [MD5/SHA-1/HMAC-SHA256/AES-CMAC OpenCL]... LWS=64 GWS=8388608 (131072 blocks) DONE
Raw:	5068K c/s real, 5178K c/s virtual

Let's use hashcat running on intel-compute runtime:
This are the demo hashes:
PMKID:

WPA*01*4d4fe7aac3a2cecab195321ceb99a7d0*fc690c158264*f4747f87f9f4*686173686361742d6573736964*** 

Messagepair:

WPA*02*024022795224bffca545276c3762686f*6466b38ec3fc*225edc49b7aa*54502d4c494e4b5f484153484341545f54455354*10e3be3b005a629e89de088d6a2fdc489db83ad4764f2d186b9cde15446e972e*0103007502010a0000000000000000000148ce2ccba9c1fda130ff2fbbfb4fd3b063d1a93920b0f7df54a5cbf787b16171000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001630140100000fac040100000fac040100000fac028000*a2

PMKID test by hashcat:

$ hashcat -m 22000 pmkid.22000 wordlist
hashcat (v7.1.2-345-g49d61d2f3) starting

OpenCL API (OpenCL 3.0 ) - Platform #1 [Intel(R) Corporation]
=============================================================
* Device #01: Intel(R) Arc(TM) A770 Graphics, 15473/15473 MB (4095 MB allocatable), 16MCU

Minimum password length supported by kernel: 8
Maximum password length supported by kernel: 63
Minimum salt length supported by kernel: 0
Maximum salt length supported by kernel: 256

Hashes: 1 digests; 1 unique digests, 1 unique salts
Bitmaps: 16 bits, 65536 entries, 0x0000ffff mask, 262144 bytes, 5/13 rotates
Rules: 1

Optimizers applied:
* Zero-Byte
* Single-Hash
* Single-Salt
* Slow-Hash-SIMD-LOOP

Watchdog: Temperature abort trigger set to 90c

Host memory allocated for this attack: 1095 MB (12689 MB free)

Dictionary cache built:
* Filename..: wordlist
* Passwords.: 718109
* Bytes.....: 8021986
* Keyspace..: 718109
* Speed.....: 573 MiB/s
* Runtime...: 0.01s

4d4fe7aac3a2cecab195321ceb99a7d0:fc690c158264:f4747f87f9f4:hashcat-essid:hashcat!
                                                          
Session..........: hashcat
Status...........: Cracked
Hash.Mode........: 22000 (WPA-PBKDF2-PMKID+EAPOL)
Hash.Target......: pmkid.22000
Time.Started.....: Sun Nov  2 09:26:18 2025 (0 secs)
Time.Estimated...: Sun Nov  2 09:26:18 2025 (0 secs)
Kernel.Feature...: Pure Kernel (password length 8-63 bytes)
Guess.Base.......: File (wordlist)
Guess.Queue......: 1/1 (100.00%)
Speed.#01........:   833.4 kH/s (10.10ms) @ Accel:18 Loops:256 Thr:512 Vec:1
Recovered........: 1/1 (100.00%) Digests (total), 1/1 (100.00%) Digests (new)
Progress.........: 147458/718109 (20.53%)
Rejected.........: 2/147458 (0.00%)
Restore.Point....: 0/718109 (0.00%)
Restore.Sub.#01..: Salt:0 Amplifier:0-1 Iteration:0-1
Candidate.Engine.: Device Generator
Candidates.#01...: 12345678 -> 70723731
Hardware.Mon.#01.: N/A

Started: Sun Nov  2 09:26:16 2025
Stopped: Sun Nov  2 09:26:20 2025

Everything is working as expected.

Messagepair test by hashcat:

$ hashcat -m 22000 messagepair.22000 wordlist
hashcat (v7.1.2-345-g49d61d2f3) starting

OpenCL API (OpenCL 3.0 ) - Platform #1 [Intel(R) Corporation]
=============================================================
* Device #01: Intel(R) Arc(TM) A770 Graphics, 15473/15473 MB (4095 MB allocatable), 16MCU

Minimum password length supported by kernel: 8
Maximum password length supported by kernel: 63
Minimum salt length supported by kernel: 0
Maximum salt length supported by kernel: 256

Hashes: 1 digests; 1 unique digests, 1 unique salts
Bitmaps: 16 bits, 65536 entries, 0x0000ffff mask, 262144 bytes, 5/13 rotates
Rules: 1

Optimizers applied:
* Zero-Byte
* Single-Hash
* Single-Salt
* Slow-Hash-SIMD-LOOP

Watchdog: Temperature abort trigger set to 90c

Host memory allocated for this attack: 1095 MB (12653 MB free)

Dictionary cache hit:
* Filename..: wordlist
* Passwords.: 718109
* Bytes.....: 8021986
* Keyspace..: 718109

024022795224bffca545276c3762686f:6466b38ec3fc:225edc49b7aa:TP-LINK_HASHCAT_TEST:hashcat!
                                                          
Session..........: hashcat
Status...........: Cracked
Hash.Mode........: 22000 (WPA-PBKDF2-PMKID+EAPOL)
Hash.Target......: messagepair.22000
Time.Started.....: Sun Nov  2 09:27:27 2025 (0 secs)
Time.Estimated...: Sun Nov  2 09:27:27 2025 (0 secs)
Kernel.Feature...: Pure Kernel (password length 8-63 bytes)
Guess.Base.......: File (wordlist)
Guess.Queue......: 1/1 (100.00%)
Speed.#01........:   798.9 kH/s (10.10ms) @ Accel:18 Loops:256 Thr:512 Vec:1
Recovered........: 1/1 (100.00%) Digests (total), 1/1 (100.00%) Digests (new)
Progress.........: 147458/718109 (20.53%)
Rejected.........: 2/147458 (0.00%)
Restore.Point....: 0/718109 (0.00%)
Restore.Sub.#01..: Salt:0 Amplifier:0-1 Iteration:0-1
Candidate.Engine.: Device Generator
Candidates.#01...: 12345678 -> 70723731
Hardware.Mon.#01.: N/A

Started: Sun Nov  2 09:27:25 2025
Stopped: Sun Nov  2 09:27:27 2025

Everything is working as expected.

Convert hashes to john's native format by hcxhashtool
PMKID:

4d4fe7aac3a2cecab195321ceb99a7d0*fc690c158264*f4747f87f9f4*686173686361742d6573736964

Messagepair:

TP-LINK_HASHCAT_TEST:$WPAPSK$TP-LINK_HASHCAT_TEST#N4OnXgDw6ZvQGPSeGAsgmub/zO2kzmyvyozHg4DFeHYUgDTTJ8L9xsSlML2Esvsv.3dWbcbS06pe9xl6bPUup5ND9FVfbBsJF4uL9U21.5I0.Ec............/GAsgmub/zO2kzmyvyozHg4DFeHYUgDTTJ8L9xsSlML2.................................................................3X.I.E..1uk2.E..1uk2.E..1uk0U...................................................................................................................................................................................../t.....U....7.6bZG79zwdIIbP1RWO4w:22-5e-dc-49-b7-aa:64-66-b3-8e-c3-fc:6466b38ec3fc::WPA2:verified:converted by hcxhashtool

PMKID test by john failed:

$ john -w:wordlist --format=wpapsk-opencl pmkid.john
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Using default input encoding: UTF-8
Loaded 1 password hash (wpapsk-opencl, WPA/WPA2/PMF/PMKID PSK [PBKDF2-SHA1 HMAC-SHA256/AES-CMAC OpenCL])
Self test failed (cmp_all(1))

Messagepair test by john failed:

$ john -w:wordlist --format=wpapsk-opencl messagepair.john
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Using default input encoding: UTF-8
Loaded 1 password hash (wpapsk-opencl, WPA/WPA2/PMF/PMKID PSK [PBKDF2-SHA1 HMAC-SHA256/AES-CMAC OpenCL])
Self test failed (cmp_all(1))

Is it now easier to understand what I mean?

I guess something changed in intel-compute-runtime and john can't handle the change.

[magnumripper]

PMKID test by john failed

Please note that the self-test isn't affected by your file content at all, so we can't be sure it's PMKID that fails. You should get the exact same failure just running john --format=wpapsk-opencl --test.

If you also bump verbosity as in john --format=wpapsk-opencl --test --verb=5, the failing test vector should be printed - please try that!

Also BTW you can add --skip-self-test (when supplying that pmkid.john file for cracking) and see if your PMKID hash does crack then.

[ZerBea]

$ john --format=wpapsk-opencl --test --verb=5
initUnicode(UNICODE, RAW/RAW)
RAW -> RAW -> RAW
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Benchmarking: wpapsk-opencl, WPA/WPA2/PMF/PMKID PSK [PBKDF2-SHA1 HMAC-SHA256/AES-CMAC OpenCL]... Loaded 8 hashes with 7 different salts to test db from test vectors
Build time: 7.140 ms
LWS=1024 GWS=1024 (1 blocks) FAILED (cmp_all(1) $WPAPSK$Coherer#..l/Uf7J..qHUXMunTE3nfbMWSwxv27Ua0XutIOrfRSuv9gOCIugIVGlosMyXdNxfBZUAYmgKqeb6GBPxLiIZr56NtWTGR/Cp5ldAk61.5I0.Ec.2...........nTE3nfbMWSwxv27Ua0XutIOrfRSuv9gOCIugIVGlosM.....)

Same hash running hashcat:

Session..........: hashcat
Status...........: Cracked
Hash.Mode........: 22000 (WPA-PBKDF2-PMKID+EAPOL)
Hash.Target......: coherer.22000
Time.Started.....: Sun Nov  2 18:15:03 2025 (1 sec)
Time.Estimated...: Sun Nov  2 18:15:04 2025 (0 secs)
Kernel.Feature...: Pure Kernel (password length 8-63 bytes)
Guess.Base.......: File (wordlist)
Guess.Queue......: 1/1 (100.00%)
Speed.#01........:   712.2 kH/s (10.10ms) @ Accel:18 Loops:256 Thr:512 Vec:1
Recovered........: 1/1 (100.00%) Digests (total), 1/1 (100.00%) Digests (new)
Progress.........: 589839/761237 (77.48%)
Rejected.........: 15/589839 (0.00%)
Restore.Point....: 442378/761237 (58.11%)
Restore.Sub.#01..: Salt:0 Amplifier:0-1 Iteration:0-1
Candidate.Engine.: Device Generator
Candidates.#01...: daniel06 -> Meriam123@
Hardware.Mon.#01.: N/A

[ZerBea]

Part 2 of the requested test (using -skip-self-test )

$ john -w:wordlist --skip-self-test --format=wpapsk-opencl coherer.john
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Using default input encoding: UTF-8
Loaded 1 password hash (wpapsk-opencl, WPA/WPA2/PMF/PMKID PSK [PBKDF2-SHA1 HMAC-SHA256/AES-CMAC OpenCL])
Note: Passwords longer than 21 [worst case UTF-8] to 63 [ASCII] rejected
Note: Minimum length forced to 8 by format
LWS=16 GWS=1048576 (65536 blocks) 
Press 'q' or Ctrl-C to abort, 'h' for help, almost any other key for status
Warning: Only 761204 candidates buffered, minimum 1048576 needed for performance.
0g 0:00:00:00 DONE (2025-11-02 18:25) 0g/s 776738p/s 776738c/s 776738C/s         ..zzzzzzzzzzzzzzz
Session completed

As expected, JtR failed.
If you have an idea what went wrong, or if you need some more tests, please let me know.

[ZerBea]

Discovered another one:

$ john --format=strip-opencl --test
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Benchmarking: strip-opencl, Password Manager [PBKDF2-SHA1 AES OpenCL]... FAILED (cmp_all(1))
[zerobeat@tux2 ~]$ john --format=strip-opencl --test --verb=5
initUnicode(UNICODE, RAW/RAW)
RAW -> RAW -> RAW
Device 1@tux2: Intel(R) Arc(TM) A770 Graphics
Benchmarking: strip-opencl, Password Manager [PBKDF2-SHA1 AES OpenCL]... Loaded 2 hashes with 2 different salts to test db from test vectors
Build time: 1.387 ms
LWS=1024 GWS=1024 (1 blocks) FAILED (cmp_all(1) $strip$*66cd7a4ff7716f7b86cf587ce18eb39518e096eb152615ada8d007d9f035c20c711e62cbde96d8c3aad2a4658497a6119addc97ed3c970580cd666f301c63ce041a1748ee5c3861ada3cd6ee75b5d68891f731b3c2e3294b08e1)

Also related to PBKDF2.

[solardiz]

I guess something changed in intel-compute-runtime and john can't handle the change.

I guess some of our OpenCL code gets miscompiled by the new intel-compute-runtime. This doesn't necessarily mean we have a bug and other projects don't - miscompiles may happen or not depending on seemingly irrelevant source code differences. We're seeing plenty of those for AMD, fewer for Intel and NVIDIA. We occasionally add workarounds. We manage to achieve no failing formats at all with recent NVIDIA drivers, but Intel and AMD are "second class" for us, so it's unfortunately expected to see some failures there.

Discovered another one:

It would actually be useful for us to see an all-formats --test=0 run with old vs. new intel-compute-runtime, but I doubt you can easily do this for the old version now? If so, at least for the new one.

[ZerBea]

I can do it (because I'm running Arch Linux), but it will take awhile.