This document reviews the public proof-of-concept material available during the disclosure window. It does not redistribute third-party exploit code; links point to the original repositories. All execution was in a disposable lab, on loopback, against official PHP images.
Scope note: the RCE proof is on the official php:8.4.26-apache image. A Nextcloud 33.0.x test
deployment (PHP 8.4.21, mod_php prefork) was used to validate the exposed configuration and
patch level; the RCE was not executed against that deployment. The bug is in PHP request
startup, not in Nextcloud code.
The lab proof-of-concept code itself is kept in a separate private, invite-only repository, available on request from the author. This document reviews and links the public PoCs; it does not redistribute them.
| Repository | What it is | Independent result |
|---|---|---|
9bie/php2shell-bypass-aslr-exp |
Python ASLR-bypass exploit, standard library only | Functional in the lab: 8/8 with ASLR on, 5/5 with ASLR off, display_errors=On; 0/3 with display_errors=Off and ASLR on (author reports 10/10 on the stock 128M/150-worker profile); host-layout dependent |
ikow/php2shell-poc |
Five chains; A/B/C public, D/E with public verification scripts and paid full exploits | A/B/C readable and testable; D/E verification public; full D/E not public |
GoldraK/php2shell-poc |
Fork/rewrite by the same author; A/C/E scanner, lab and rules | Scanner reports 24/24 vulnerable; measures A/C/E, not the OOM chain |
dinosn/php-cve-2026-10961x-lab-poc |
Lab with instrumented allocator | Lab-assisted; uses USE_ZEND_ALLOC=0; not a universal remote PoC |
pentestershell/php2shellbt |
Small scanner | Confirmed false positive on 8.4.27; CVEs mislabeled |
eness/php2shell-...-PoC |
Placeholder | No code published |
KillaBoi/StealMyBitcoin-PoC |
End-to-end 4-stage exploit resolving the stealmybitcoin challenge | Confirmed 100% in lab: resolves Chain D (forged HashTable -> zif_system) against php -S with workers |
- Python 3, standard library only, 20 HTTP requests maximum.
- SAPI Architecture Disambiguation: Operates against Alpine
PHP 8.4.26running the built-in CLI server (php -S 0.0.0.0:8000) with multiple workers (PHP_CLI_SERVER_WORKERS=4) behind an Nginx reverse proxy (proxy_passHTTP).- Crucial distinction: CLI-server workers persist across OOMs (identical to
mod_php), allowing the 256-byte cookie leak to return Zend heap pointers andzval_ptr_dtor. - In standard PHP-FPM, worker aborts (
exit 70) prevent this leak mechanism, confirming our findings in section 11.
- Crucial distinction: CLI-server workers persist across OOMs (identical to
- Exploitation Mechanism (Chain D):
- Stage 1 (Heap Leak): 1 producer (249 KB multipart OOM) + 8 pipelined consumers -> extracts 256 B cookie disclosing aligned 2 MiB Zend MM heap base.
- Stage 2 (Code Leak): 1 producer (251 KB multipart OOM) + 8 pipelined consumers -> extracts 16 B cookie disclosing
zval_ptr_dtor(0x112930offset), calculatingzif_system = zval_ptr_dtor - 0x204BC0. - Stage 3 (Callback): 1 POST (368 KB) deploying a 384-byte forged
HashTableoverlapping a fake call frame andzend_stringcommand. PHP request cleanup readspDestructoraszif_systemand executes the command. - Exfiltration: Scans
/proc/self/fd/to write stdout directly back to the active accepted HTTP client socket.
- Independent Lab Verification: Replicated in our
cachyenvironment on 2026-10-26 againstphp:8.4.26-fpm-alpine3.23+ Nginx; achieved full command execution (UID 0) and recovered the test flag.
- Python 3, standard library, no external hosts.
- The exploit targets the OOM/state-corruption chain and includes an ASLR bypass.
- In our lab it produced
uid=33(www-data)on 8/8 runs with ASLR on and 5/5 runs with ASLR off (raw:raw/9bie-aslr-on-8run.txt; the 8 ASLR-on runs used 8 distinct chunk andlibphpbases;raw/9bie-aslr-off-5run.txt; one base across all 5 off runs). In the same lab profile withdisplay_errors=Offand ASLR on it stops at stage 1 in 0/3 runs (raw/9bie-display-off-3run.txt); the author's 128M/150-worker Off profile was not reproduced here. - The earlier 0/49 "Off" sweep was a harness bug: stage A was sent to the wrong port while stage B hit the container port. It was not a PoC failure.
- The author's published run uses the official profile (
php.ini-production,display_errors=Off,memory_limit=128M,MaxRequestWorkers=150) and reports 10/10 on x86_64 and aarch64. Our 8/8 and 5/5 runs used the lab profile (single worker,memory_limit=8M,display_errors=On); the 150-worker stock profile was not re-run here. - Caveat: the chain depends on host/kernel allocator layout. The same public exploit did not reproduce on every host tested; the exploit is not universal without calibration.
- A: session upload-progress filename injection + LFI.
- B: local
/proc/self/memprimitive; not remote. - C: session handler mismatch across directories, plus an application gadget.
- D/E: OOM/state chains. The repository publishes verification scripts; the full exploits are not public in the reviewed state.
- The README's claims are useful but are claims; each chain needs its own prerequisites and
proof. Our independent tests are in
EVIDENCE.md.
- Fork of the previous repository, rewritten by the same author (12 commits ahead, 6 behind at review time).
- Scanner covers A/C/E against a long version list; it reports 24/24 "vulnerable".
- The lab configuration is intentionally weak:
session.use_only_cookies=0,session.use_strict_mode=0,session.upload_progress.cleanup=Off. That makes A easy but does not represent a hardened default. - The scanner does not measure the OOM/state-corruption chain. 8.4.27 appears in
versions.jsonbut not in the 24-row result table. - Our differential: A/C/E work on 8.4.26, 8.4.27 and 8.5.12 with the lab configuration. With
use_only_cookies=1,strict=0and SID only in the multipart body, A still works;strict=1blocks it. C is per-directory handler mismatch plus gadget; E requires local file planting in the harness. None of these results contradicts the 8.4.27 fix for the OOM chain. - A separate lab dropper based on the same upload-progress write can keep a shell-readable file
alive during a slow upload even with
cleanup=On; execution still requires a command primitive such assh</tmp/sess_<id>. It is a post-exploitation write primitive, not a standalone remote RCE.
- Reproduces the bug with an instrumented allocator and
USE_ZEND_ALLOC=0. - The repository itself scopes the result as lab-assisted; it is not a turnkey remote exploit.
- Labels CVE-2026-109618 as "session fixation" and CVE-2026-109619 as "var_destroy UAF"; neither mapping matches the advisories.
- Its check 2 marks 8.4.27 vulnerable because it only looks for session warnings with
display_errors=On. - Its check 1 ignores PHP 7.x, 8.0 and 8.1.
- It disables TLS verification (
ssl.CERT_NONE). - Verdict: not a reliable detector; do not use it as patch evidence.
No code was published at review time.
A useful PoC should state:
- the exact PHP version and SAPI (mod_php/FPM, prefork/worker);
- the relevant
php.inivalues (memory_limit,post_max_size,session.upload_progress.*,display_errors,session.use_strict_mode); - whether ASLR is on or off, and how the leak works;
- the network/proxy conditions (buffering, KeepAlive, worker count);
- the proof (command output,
uid, file write) and the failure modes; - that it is authorized lab use, with no hardcoded external targets or disabled TLS verification.
A scanner that only observes warnings is not proof of RCE. A CVE identifier is not proof of exploitability on a given build.
See EVIDENCE.md for the full matrices. The short version:
- 8.4.26
mod_php,display_errors=On, ASLR on: RCE 3/3. - 8.4.26
mod_php,display_errors=Off, ASLR off: RCE 3/3. - 8.4.26
mod_php,display_errors=Off, ASLR on: our cookie-only chain gave controlled repetitions 0/3; the 4000-byte cookie is state-dependent and over-reads into heap objects (47 pointer-like values per run); no session-table read, and the chunk-to-libphp distance varies 1.88 MB (fixed page offset, ~481 pages). The 9bie chain reaches RCE on the same build with ASLR on (see the 9bie entry above). - Chain D (mod_php stale EG + forged
zend_objectspray ->zif_system): described inikow/php2shell-pocdocs/ANALYSIS.md(sections 3.5/3.6); the exploit scripts were not public at review time. We independently confirmed its OOM precondition (chunkedpost_max_sizebypass on unpatchedmod_php) and then inspectedEG()atphp_request_shutdownon the same worker:EG(objects_store)(top=1 size=1024 flh=-1),vm_stackandcurrent_execute_dataare identical for the healthy request, the OOM request and the next request, with no dangling pointer or crash. The stale-EG step is therefore not supported by our measurements; the OOM precondition is. - CVE-2026-115560: bounded reproduction attempt (config
php_binary+ runtimeini_set()tophpbeforesession_start(), upload progress on). The condition reproduces a decode failure (Session object destruction failedfamily) but no crash or memory-safety symptom in our runs; inconclusive until the full advisory is public. - Community backports for EOL PHP 5.4-8.1 exist (e.g.
Anton-Barinov/php2shell-patches); they are third-party code we did not review for correctness. - 8.4.27: 0/5, no aborts or over-read.
- A/C/E: RCE with lab prerequisites on 8.4.26, 8.4.27 and 8.5.12.
- A with
strict=1: blocked in the tested configuration. PS(serializer): no stale handler across requests in FPM ormod_php.- E-OOM (FPM stale
BG(unserialize)): deterministic UAF crash (502) on 8.4.26 with a symbol build; with the multipart-reclaimed struct (planted 24-byte session, ASLR off) the fakevar_dtor_entriesgives a controlled 16-byte write tosapi_module.ub_write=&systemand command execution in the lab; 8.4.27 stable; the remote ASLR-on variant was not demonstrated.