CVE-2026-109618 PHP2Shell exploitation detected
Goal
Detect exploitation of CVE-2026-109618 and CVE-2026-109619 (PHP2Shell), PHP multipart and session upload-progress memory corruption vulnerabilities that enable remote code execution.
Strategy
Exploitation attempts crash PHP workers, and successful exploitation runs attacker code inside a worker. The request payloads are filler, so the primary rules watch the worker process instead of the network. A supplementary packet canary flags the exploit payload itself on cleartext ingress.
- Critical: A PHP worker (
php-fpm, php-cgi, lsphp) spawned a shell, downloader, interpreter, or memfd binary within 15 minutes of a worker crash in the same container. - High: A PHP worker spawned one of those programs with no recent worker crash.
- Medium: A PHP worker exited with code 70 (the pre-request memory-limit fatal of the arm stage) or crashed with SIGSEGV (the fire stage). Every qualifying crash refreshes the 15-minute Critical correlation window; the backend deduplicates the Medium signal per container.
- Medium (stream): The PHP2Shell exploit payload was observed on the wire — a large packet whose payload begins with the
Z-filler run (0x5A5A5A5A) or the forged HT_MIN_MASK table mask (0xFEFFFFFF). Fires once per host per 5 minutes.
sendmail invocations from PHP mail() are excluded. Some applications legitimately run shell commands from PHP; baseline the High case before alerting on it.
The stream match is a best-effort canary, not the backbone. It only sees cleartext IPv4 traffic at the host and keys on the public PoC’s byte shape, so TLS-terminated ingress, IPv6, reshaped payloads, or bytes that do not land at the start of a captured segment all evade it (the capture window is 256 bytes per packet, and packet.filter is a fixed-offset BPF match with no content search). Its process attribution is the sending side (for example an ingress proxy or docker-proxy), not the targeted PHP worker, so it correlates host-wide and by time only — it does not share the cgroup-scoped correlation of the crash and exec legs. Treat it as “exploit attempt seen”; rely on the worker legs for confirmation of success.
Triage and response
- Identify the workload from
@container.id and host, and the program the worker ran from @process.executable.path and @process.args. - For Critical and High signals, look for outbound connections, downloads, or memfd executions from the same container.
- For Medium signals, check the web server access log for large multipart POSTs from one source around the crash, followed by 500 or 502 responses.
- For stream Medium signals, map the destination port to the service and correlate with any worker crash or exec on that host; a stream signal with no worker crash is likely a blocked or failed attempt.
- Confirm the PHP version is below 8.5.12, 8.4.27, 8.3.36, or 8.2.35.
- Isolate the host or terminate the container, then patch PHP and restart PHP-FPM or CGI workers.
- Until patched, set
session.upload_progress.enabled=Off.
Requires Agent version 7.68 or greater.