Documentation

CVE coverage

Pahlevan does not detect vulnerabilities. It observes what a workload does, and it refuses what the workload has not been seen to do before. This page is about the gap between those two sentences and a headline CVE, because that gap is where security tools overclaim.

Each row below names a vulnerability that was exploited or published recently, says what Pahlevan would see after it was exploited, names the kernel hook that sees it, says whether that is observation or enforcement, and states the limitation. Where a row's answer is "nothing", the row says so.

Three distinctions run through the whole page, and they are the only reason it is worth reading:

  1. The initial code execution is never observable. It happens inside the vulnerable process. A memory-corruption bug in a PHP worker produces no exec, no file open and no outbound connection: the worker is already running and already has its sockets. Pahlevan's hooks fire on syscall-boundary events, so there is nothing for them to fire on. No row below claims otherwise.
  2. A targeted detection and a baseline deny are different claims. A targeted detection fires on the shape of the event itself, whatever the workload learned. A baseline deny fires because the operation was not in the learned allow-set, and would not fire in a workload that had learned it. Each row says which one it is. A web shell that is caught only because sh was never in the nginx baseline is written here as a baseline deny, not as a signature for the CVE that planted it.
  3. Ingress is not observed at all. The network hook is lsm/socket_connect, which is egress only. The inbound request that carries the exploit is not seen by anything in Pahlevan, on any kernel.

How these were verified

Every CVE here was checked against the NVD API (services.nvd.nist.gov), whose record carries CISA's cisaExploitAdd date for any CVE in the Known Exploited Vulnerabilities catalog, and against the vendor's own advisory where one exists. The KEV dates and the catalog names quoted below are those fields. The cisaExploitAdd date, the cisaActionDue date and the CVSS vectors are quoted from the record rather than from reporting about it, because the figures in circulation after a leak are frequently not the vendor's own.

Two CVEs below are not in the KEV catalog, and the page says so where they appear rather than letting a reader assume that a page of KEV entries is uniformly KEV.

PHP2Shell: CVE-2026-109618 and CVE-2026-109619

Two flaws in PHP itself, patched in an emergency release at 14:00 UTC on 2026-10-21 after details leaked ahead of coordinated disclosure. Both are rated CVSS v4.0 9.2 by php.net; NVD's own CVSS v3.1 base scores are 9.8 and 8.1. A CVSS 10.0 figure circulating from the leak is not PHP's rating and is not used here. Neither CVE is in the KEV catalog as of 2026-10-25.

Patched in PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35.

The affected range is disputed. The php-src advisories list only the supported branches, as versions below each of those four releases. Wiz, one of the discoverers, says PHP 5.4 through 8. watchTowr notes that no official vendor advisory confirms that wider scope. This page does not resolve the dispute and does not pick a side.

Exploitation status is disputed. watchTowr reported in-the-wild exploitation on 21 October. Huntress reported none on partner systems. Cloudflare saw early attempts with no confirmed breaches. The absence of either CVE from the KEV catalog is consistent with the dispute being unresolved rather than with either claim being settled.

CVE-2026-109618 is an out-of-memory path that longjmps out of a context holding partly initialised request memory, which the cleanup then reads or branches on, reaching remote code execution as the web worker account. The advisory confirms this against mod_php and says FPM, FrankenPHP and LSAPI are susceptible on source analysis, with no complete FPM exploit demonstrated.

CVE-2026-109619 is stale global state still pointing into freed request memory. The advisory describes the confirmed impact as unauthenticated heap use-after-free, memory disclosure and remote code execution with the web worker's privileges.

The memory disclosure is a complete blind spot for Pahlevan, and it is worth stating on its own:

A request that reads PHP's own heap and returns the bytes in a response performs no syscall that any of Pahlevan's hooks attach to. There is no file open, no exec, no connect, no credential change and no capability check. The process reads memory it already owns. Pahlevan sees nothing, reports nothing, and cannot be configured to. A reader who wants that class of attack covered needs a different kind of tool, and Armis's own mitigation for it is to restart workers so the abused in-memory state is cleared, which is an action Pahlevan plays no part in.

What is observable is what an attacker does with the code execution once they have it, and that is the rest of this page.

Sources: https://github.com/php/php-src/security/advisories/GHSA-w8hw-w9fx-j79w, https://github.com/php/php-src/security/advisories/GHSA-6wmf-4rv5-p3rp, https://research.jfrog.com/post/php2shell-all-you-need-to-know-about-cve-2026-109618-and-cve-2026-109619/, https://watchtowr.com/intelligence/php-cve-2026-109618-faq/, https://www.armis.com/blog/php2shell-how-to-find-prioritize-and-patch-affected-systems/, https://www.helpnetsecurity.com/2026/10/21/php2shell-php-vulnerabilities-patched-cve-2026-109618-cve-2026-109619/

The matrix

Kind is the distinction that matters: targeted means the detection keys on the structure of the event and fires whatever the workload learned; baseline means it fires only because the operation was outside the learned allow-set, and a workload that had learned the same operation would not report it.

Mode is observation against enforcement. Enforcement means the kernel refused the operation before it took effect, which requires the BPF LSM (lsm=bpf on the kernel command line; see the BPF LSM). Everything marked observation is a report after the fact, including both targeted signals: they are computed on the export path in userspace, after the kernel has already allowed or denied the operation, so a signal never blocks anything.

CVE / technique Post-exploitation step Pahlevan sees Hook Kind Mode Honest limitation
CVE-2026-109618, CVE-2026-109619 (PHP) The initial RCE: nothing. none - - Happens inside the php-fpm or mod_php worker. No syscall boundary is crossed, so no hook fires.
CVE-2026-109619 (PHP), memory disclosure Nothing. none - - Reads the process's own heap. Produces no kernel event of any kind. A blind spot, not a gap in configuration.
CVE-2026-109618, CVE-2026-109619 (PHP), command execution after RCE A shell or interpreter execed with a request-serving process in its ancestry, for example php-fpm -> sh. lsm/bprm_check_security targeted (server-spawned-shell) observation Keys on the parent's comm, which is attacker-influenced: a payload that renames itself, or that never execs at all and does its work through PHP's own functions, is not seen. Enforcement of the same exec is a baseline deny, below.
CVE-2026-109618, CVE-2026-109619 (PHP), any unlearned exec The exec itself, refused if it was never learned. lsm/bprm_check_security baseline enforcement Only as good as the baseline. An image whose entrypoint is a shell wrapper has learned sh, and then this does not fire at all; that is exactly why the targeted signal above exists.
CVE-2026-12569 (PTC Windchill and FlexPLM) A .jsp written into a served directory, which is how this campaign was found: hex-named JSP files under the Windchill login directory. lsm/file_open targeted (web-shell-write) observation The directory list is a list; a deployment serving from a path not in it is missed. A CI job deploying a .php into a document root looks identical, which is why this reports rather than blocks.
CVE-2026-12569 (PTC Windchill and FlexPLM) The same write, refused if the path was never learned as writable. lsm/file_open baseline enforcement A workload that writes to its own document root during learning, for a cache or an upload directory, grants that write permanently.
CVE-2026-12569 (PTC Windchill and FlexPLM) The JVM running a command after the shell is fetched: java -> sh. lsm/bprm_check_security targeted (server-spawned-shell) observation java is a request-serving runtime here and a build tool elsewhere; a JVM workload that legitimately execs shells will report this and should be tuned or excluded.
CVE-2026-104935 (Ivanti Connect Secure) The arbitrary commands the auth bypass allows, as execs outside the learned set. lsm/bprm_check_security baseline enforcement Scope: Connect Secure is an appliance. Pahlevan sees it only if it runs as a workload on a node Pahlevan instruments, which is not how it is normally deployed. The row is honest about the shape, not a claim of coverage.
CVE-2026-102764 (BeyondTrust Remote Support) The file operations the flaw allows, as writes outside the learned set. lsm/file_open baseline enforcement Same appliance scope caveat. Reporting also places the post-exploitation log deletion attributed to this campaign within reach of the file hook, again as a baseline deny.
CVE-2026-96656 (Plex Media Server) A shared object written where the server loads modules from, to be executed on the next restart. lsm/file_open baseline enforcement Baseline only. There is no targeted signature for a module drop: the directory is server-specific and a legitimate plugin install is the same write. A workload that installs plugins at runtime has learned it.
CVE-2026-96652, -96654, -96655, -96656 (Plex2Shell), command execution Plex Media Serv -> sh, the truncated comm the kernel reports for Plex Media Server. lsm/bprm_check_security targeted (server-spawned-shell) observation Matches one exact truncated comm. A differently named build is missed.
Plex2Shell payload: XMRig miner The miner's own exec, and its egress to a pool address. lsm/bprm_check_security, lsm/socket_connect baseline enforcement Both are baseline denies and nothing more: Pahlevan has no miner signature, no pool address list and no hash-rate heuristic. A pool reachable at an address the workload already connects to is not distinguished.
Plex2Shell payload: proxy malware The proxy's exec and its egress. lsm/bprm_check_security, lsm/socket_connect baseline enforcement As above. Inbound proxy connections are not seen at all: the hook is socket_connect, egress only.
Service account token read after any of the above The open of the projected token path. lsm/file_open baseline enforcement A workload that legitimately reads its own token, which most that talk to the API server do, has learned it.
Credential change after any of the above Every commit_creds, discriminated by whether an execve was in progress. kprobe/commit_creds baseline observation A kprobe cannot refuse anything. Observation only on every kernel, with or without the BPF LSM.

Verified, in scope of nothing

These were verified as exploited and are deliberately not given a coverage row, because an honest answer is that Pahlevan cannot see them:

  • CVE-2026-0322 (Palo Alto PAN-OS). KEV, added 2026-10-14, due 2026-10-17, catalog name "Palo Alto Networks PAN-OS Code Injection Vulnerability". Authenticated RCE as root on the management interface of PA-Series and VM-Series firewalls and Panorama. PAN-OS is a firewall, not a Kubernetes workload. Nothing Pahlevan does applies to it, and a coverage row would be a claim about a product it is never installed on. https://security.paloaltonetworks.com/CVE-2026-0322

Nothing was dropped for lack of confirmation

Every candidate investigated was confirmed. For the record, each one's authoritative status:

CVE KEV cisaExploitAdd Catalog name or vendor advisory
CVE-2026-109618 not in KEV php-src GHSA-w8hw-w9fx-j79w
CVE-2026-109619 not in KEV php-src GHSA-6wmf-4rv5-p3rp
CVE-2026-104935 2026-10-13 Ivanti Connect Secure Authentication Bypass by Spoofing Vulnerability
CVE-2026-0322 2026-10-14 Palo Alto Networks PAN-OS Code Injection Vulnerability
CVE-2026-102764 2026-10-16 BeyondTrust Remote Support Authentication Bypass by Spoofing Vulnerability
CVE-2026-96652 2026-10-20 Plex Media Server Server-Side Request Forgery (SSRF) Vulnerability
CVE-2026-96654 2026-10-20 Plex Media Server Improper Neutralization of Encoded URI Schemes in a Web Page Vulnerability
CVE-2026-96655 2026-10-20 Plex Media Server Server-Side Request Forgery (SSRF) Vulnerability
CVE-2026-96656 2026-10-20 Plex Media Server External Control of File Name or Path Vulnerability
CVE-2026-12569 2026-06-25 PTC Windchill and FlexPLM Improper Input Validation Vulnerability

Sources: https://www.cisa.gov/known-exploited-vulnerabilities-catalog, https://hub.ivanti.com/s/article/Security-Advisory-Ivanti-Connect-Secure-CVE-2026-104935, https://www.beyondtrust.com/blog/entry/bt26-05, https://www.ptc.com/en/support/article/CS473270, https://zmain.info/blog/plex2shell, https://thehackernews.com/2026/10/hackers-exploit-plex2shell-flaws-to.html

The two targeted signals, precisely

Both are implemented in pkg/export/postex.go and appear on the exported envelope as signals, on an OTLP record as pahlevan.signals, and in the pahlevan_postex_signals_total metric.

server-spawned-shell

Fires when an exec is of a shell or a scripting interpreter and a request-serving process appears in its ancestry. Both halves are necessary and neither is sufficient.

Why the ancestry and not the immediate parent: the second command an attacker runs is a grandchild of the server, so php-fpm -> sh -> python3 is the same intrusion as php-fpm -> sh, and a rule reading only the parent would report the first and miss every one after it.

Why this is not a generic deny: the exec of sh is unremarkable. In an image whose entrypoint is a shell wrapper it is in the learned baseline, and the baseline deny has nothing to say about it. The parent is the entire signal.

What it does not catch:

  • A payload that does its work inside the interpreter, through PHP's own file and network functions, and never execs anything.
  • A server whose comm is not in the list in postex.go. The list is explicit and finite.
  • A shell execed by a process the attacker has already renamed.

Known false positives: a JVM or Node workload that legitimately execs shells, and any application server with a shell-based health check or an init hook that runs after the server itself has started.

web-shell-write

Fires when a write-mode open names a file with a server-side script extension and the path lies under a directory a web server serves. Both halves are necessary: a .php written to /tmp is staging and a .png written to a document root is an upload, while a .php in a document root is remotely reachable code the moment it lands.

The extension is also checked after stripping one trailing suffix, because the standard bypass is to land the file as shell.php.bak and rename it: the write is the observable event and it happens under the suffixed name.

What it does not catch:

  • A deployment serving from a directory not in the list.
  • A script with an extension not in the list, including anything the server has been configured to execute through a non-standard handler.
  • The rename itself. The hook is lsm/file_open; a rename(2) that turns an already-written file into a reachable script is seen by the syscall tracepoint and not by this signal.

Known false positive: any deployment that writes its own application code at runtime, which a CI job deploying into a document root does by design.

Benchmark scenarios for these shapes

test/benchmark/scenarios/ models each shape, so a claim on this page can be run rather than read:

Scenario Shape Relevant rows above
27-server-spawns-shell A process whose comm is a request-serving runtime execs a shell. PHP2Shell command execution, CVE-2026-12569, Plex2Shell command execution
28-webshell-file-drop A server-side script planted in a served directory. CVE-2026-12569
29-server-module-drop A shared object written where a server loads modules from. CVE-2026-96656

Scenario 27 runs in an nginx pod and so copies a shell to a path named php-fpm before running the inner shell from it. What the kernel reports is a process whose comm is php-fpm execing sh, which is the lineage a real FPM worker produces; installing PHP would not make those two process records any more real. The scenario header says so, and the modelling choice is stated there rather than left for a reader to discover.