On November 19, 2026, PHP shipped fixes for phpAfterShell and 16 other vulnerabilities across 8.2.36, 8.3.37, 8.4.28, 8.5.13 and 8.6.0. Not every release carries the full set: 8.5.13 and 8.6.0 fix all 17, while 8.4.28 fixes 14 and 8.2.36 and 8.3.37 fix 13 each where branches are unaffected, per Precursor’s tracker. For PHP-FPM operators, the practical answer is to install the appropriate fixed release and restart every worker. A WAF can reduce exposure during that rollout, but it cannot complete it. Cloudflare’s vendor-reported guidance is unusually direct: “The WAF is an additional layer of defense, not a replacement for the patch.”
There is also a configuration trap immediately after upgrading. The fixed releases turn session.upload_progress.enabled off by default, but an explicit session.upload_progress.enabled=1 in php.ini, an FPM pool or a virtual host overrides that safer default. A fleet can therefore run the patched binary while retaining the upload-progress entry path. The version serving web requests and the effective configuration both need checking.
Map each stopgap to the attack-path stage it interrupts
An interim control is useful only when its coverage is stated precisely. Blocking a suspicious HTTP request, disabling the upload-progress callback and changing how the front end connects to FPM interrupt different stages. None repairs every affected PHP component.
| Interim control | Stage it interrupts | What it does not establish | Operational use |
|---|---|---|---|
| Cloudflare’s updated PHP RCE rule | Inbound HTTP requests matching known phpAfterShell patterns. Cloudflare reports blocking on all plans, including Free, through a rule updated on November 19. | This is vendor-reported coverage, not independent verification that every exploit variant is blocked. It does not repair the session handler or origin-side FPM behavior. | Useful during the rollout, followed by the PHP upgrade and worker restart. |
Cloudflare rules for CVE-2026-129289, CVE-2026-15353 and CVE-2026-129506 | Request detection for the FPM and SOAP patterns. These rules default to Log, not Block. | Logging alone does not stop a matching request. Cloudflare also says some critical flaws use paths its WAF does not see. | Review Security Events and switch to Block only after confirming that legitimate traffic is unaffected. |
session.upload_progress.enabled=0 | The RFC 1867 upload-progress entry path into the vulnerable file-based session handler. | It does not repair the handler. Applications that place attacker-controlled data into sessions through other routes remain affected. | A stopgap for hosts that cannot yet be restarted, not a reason to defer the patch. |
| Closing front-end FastCGI connections after each request | A required precondition of the demonstrated CVE-2026-129289 request-smuggling attack. | It says nothing about the other 16 CVEs, and reporting conflicts over how common the vulnerable configuration is. | Check nginx, Apache, HAProxy or LSAPI settings explicitly; do not infer safety from using a mainstream front end. |
| Redis, database or another session store that replaces records in one step | The file handler’s interrupted, non-atomic session-file update. | It does not remove exposure to the rest of the November release. PHP also advises checking how each store handles an interrupted update. | Treat it as a narrower exposure statement for the file-handler flaw, not a fleet-wide exemption. |
The WAF logging distinction matters. According to Cloudflare’s post, the three rules ship with Log as their default because their request patterns can also occur in legitimate traffic. That is a deliberate false-positive decision, not evidence that the requests are harmless. Until an operator reviews the events and enables blocking, these rules provide telemetry rather than an active barrier.
The upload-progress stopgap has a similarly bounded effect. The Hacker News’s account of the maintainer guidance states: “Switching the feature off is not a full fix. It closes the upload path, but apps that put attacker-controlled data into sessions in other ways remain affected, the maintainers say.” The community proof of concept reaches the same conclusion from the other direction: its mitigated target loses the upload-progress entry path, but the session handler itself remains unrepaired.
Cloudflare’s runbook also lists fatal_error_backtraces=0 on PHP 8.5 as interim hardening for the fatal-error-backtrace use-after-free flaws, and the post names CVE-2026-129474 and CVE-2026-129182 among the server-state CVEs its configuration changes address. For the demonstrated CVE-2026-129474 attack, though, Cloudflare credits a different setting: zend.exception_ignore_args=1, the php.ini-production default. No reviewed source demonstrates fatal_error_backtraces=0 blocking an exploit, so it should not be counted as demonstrated coverage when deciding whether a host still needs patching.
The branch-to-fixed-release matrix
The November 19 releases fix 17 CVEs. Precursor Security’s tracker identifies seven as critical, each scored CVSS 4.0 9.2: CVE-2026-129289, CVE-2026-126714, CVE-2026-115560, CVE-2026-129496, CVE-2026-129474, CVE-2026-129481 and CVE-2026-129182. phpAfterShell is CVE-2026-126714, located in PHP’s default file-based session handler.
| Installed branch | Fixed release to target | Branch consequence |
|---|---|---|
| PHP 8.2 before 8.2.36 | 8.2.36 | PHP 8.2 reaches end of life on December 31, 2026, so this is also a prompt to plan the branch migration. |
| PHP 8.3 before 8.3.37 | 8.3.37 | Take the point release rather than combining this response with an unnecessary branch jump. |
| PHP 8.4 before 8.4.28 | 8.4.28 | Debian 13 has a fixed distribution build; check the package revision rather than upstream version text alone. |
| PHP 8.5 before 8.5.13 | 8.5.13 | Review any interim 8.5 hardening as temporary, including the backtrace setting above. |
| PHP 8.6 release candidates | 8.6.0 | The final 8.6.0 release shipped the same day with the same 17 security fixes as 8.5.13. |
| PHP 8.1 and older | No fix planned | These branches are end of life. Compensating controls do not put them back on a supported patch path. |
The fixed-version mapping is reported by both Cloudflare and Precursor; distribution status requires a separate check.
One recent patching milestone does not help here. Cycode’s analysis says October’s PHP2Shell releases do not contain the phpAfterShell fix. Hosts patched for PHP2Shell remain affected, including hosts that disabled upload progress during that earlier response.
Why the session flaw escapes request filtering
The community PoC describes a non-atomic session-file update using pwrite() and ftruncate(), triggered through the upload-progress callback. Its demonstrated chain interrupts that update through an out-of-memory kill, then relies on a destructor gadget to turn the resulting condition into code execution. An edge WAF may inspect the originating HTTP request, but it does not mediate the worker’s memory boundary, the mid-write interruption or the subsequent server-side gadget execution.
That demonstration also limits what can responsibly be claimed about production exploitability. The PoC’s documented lab used:
- the default
session.save_handler=files; - a memory-bounded Docker target with a 64 MiB cgroup limit;
- a bundled, representative
FileCachedestructor gadget; - low latency to the target.
It succeeded in that localhost lab. It did not establish that every production PHP host has a usable gadget, favorable timing or the same memory boundary. Conversely, the absence of a known application gadget is not a dependable safety argument. The Hacker News reports that depthfirst’s own chain needed no application gadget because it used a second bug: memory corruption in a PHP built-in class. Details are withheld, and that second bug may not be fixed in the November 19 releases.
This is the strongest reason not to turn the PoC’s preconditions into a checklist for declaring hosts safe. Those preconditions describe one reproduced chain, not the full set of possible chains.
FastCGI reuse is an origin configuration problem
The PHP-FPM request-smuggling flaw, CVE-2026-129289, makes the WAF boundary even clearer. Precursor’s reading of the advisory lists these front-end conditions:
- nginx with
fastcgi_keep_conn on; - Apache
mod_proxy_fcgiwithenablereuse=on; - HAProxy FastCGI with kept connections, or LSAPI;
- an application path that opens
php://inputwithout reading it to the end, something frameworks can do before routing.
Precursor adds: “The advisory says standard nginx and Apache configurations that close the FastCGI connection after each request are not affected by the demonstrated attack.”
ThreatAft’s assessment is broader. It calls this the broadest-reach CVE and says the requirements cover “many standard nginx and Apache deployments with framework-based applications.” Those statements cannot be reconciled from the fetched material: one limits the demonstrated attack to reused connections, while the other characterizes that deployment pattern as common. The safe operational response is to inspect connection reuse and application behavior rather than assume either that nginx is vulnerable by default or that a standard-looking configuration is unaffected.
Either way, the decisive condition sits between the front end and FPM. A rule evaluating requests at an edge WAF cannot determine whether an origin reuses its FastCGI connection.
Independent research also explains why request signatures should not be treated as complete coverage. A paper on WAF signatures describes their basic assumption: a payload contains a stable substring unique enough to identify attacks without false rejections. Cloudflare’s Log-only rationale is a concrete example of what happens when that uniqueness is weak.
In a separate, author-reported XSS experiment, GenXSS generated 220 payloads with GPT-4o; 80% bypassed the tested state-of-the-art ModSecurity rules, and the authors added 15 rules to block 83% of them (GenXSS). This was not a test of Cloudflare or phpAfterShell, so it does not demonstrate a bypass of the new PHP rules. It does support the narrower judgment that signature coverage can require iteration as payloads change.
Triage exposure without creating patch exemptions
The evidence supports differences in rollout priority, not exemptions from upgrading:
- Prioritize hosts using file-based sessions with upload progress enabled, especially memory-bounded, internet-reachable workers. These match more of the reproduced chain’s conditions.
- Inspect FastCGI reuse directly. Closing connections per request removes a precondition of the demonstrated
CVE-2026-129289attack, but not the need for the rest of the security release. - Redis and database-backed sessions sit outside the file-handler flaw when they replace records in one step, according to Cycode. Those hosts still need the other fixes and a check of interrupted-update behavior.
- Do not use the lack of a known application gadget as a clearance. depthfirst’s undisclosed second bug removes that requirement in the researchers’ own reported chain.
As of November 20, none of the seven critical vulnerabilities was listed in CISA’s Known Exploited Vulnerabilities catalog, and no reviewed source reported exploitation in the wild, according to ThreatAft. That makes this a severity-and-reachability response, not a response to documented active attacks. It does not supply evidence that waiting is safe.
The patch runbook that actually finishes the job
Cloudflare’s vendor runbook supplies the essential sequence:
- Select the fixed build for the existing branch. For most 8.2 through 8.5 fleets, that means the point release rather than an emergency migration to 8.6.0.
- Rebuild container images and deploy the fixed package.
- Restart the PHP-FPM master or Apache so every worker loads the new binary. Replacing files on disk without completing this step leaves the rollout unfinished.
- Verify the version serving web requests. A CLI binary, image tag or package-manager transaction is not enough when distribution backports and long-running workers are involved.
- Search
php.ini, FPM pool definitions and virtual hosts forsession.upload_progress.enabled=1. Remove the override where the feature is unnecessary. - Keep interim controls in place during the rollout and review WAF logs. A move from Log to Block should follow a check for legitimate matches, as Cloudflare recommends.
The upload-progress default change cuts both ways. The Hacker News reports that it was enabled in 8.2.35, 8.3.36, 8.4.27 and 8.5.12, but is off by default in the fixed releases. Applications that genuinely rely on upload progress may need to re-enable it. Those applications should then be treated as lacking that particular stopgap; the binary upgrade and complete worker restart become even more important.
Fleet realities: backports and the 8.2 clock
Distribution packages can contain the fix without changing the upstream-looking version number. As of November 19, Cycode’s distribution tracker listed Debian 12 and 13 as fixed through 8.2.36-1~deb12u1 and 8.4.28-1~deb13u1, with Debian testing and unstable still unfixed and no Alpine advisory yet. By November 20 the picture had moved: Precursor’s live tracker reports Ubuntu shipping fixes for CVE-2026-126714 and CVE-2026-129289 on all supported releases, Amazon Linux 2023 shipping all four fixed branches, SUSE shipping fixes for some SLES 15 releases, and Red Hat still without fixed packages, while CVE-2026-115560 remained awaiting triage on Ubuntu. An Ubuntu fleet may therefore already carry the two headline fixes through a routine package update and still be exposed to the rest of the release. These statuses are time-sensitive, so the vendor advisory and package revision are the evidence to check, not php -v alone.
For PHP 8.2 estates, this point release and the branch-migration plan are separate pieces of work. ThreatAft dates 8.2’s end of life to December 31, 2026. Precursor likewise presents the current-branch point release as the faster, lower-risk response for most teams, while noting that 8.6.0 contains the same fixes.
A rushed migration has its own regression risk, although the available evidence cannot quantify that risk for a particular application. In the author-reported, single-file study “When Modernization Breaks Working Code”, LLM tools fixing 6–9 files also broke 5–7 when changes were required; Rector fixed six and broke one. This is not a PHP 8.6 compatibility benchmark. It supports a narrower decision: install the security point release now, then test and schedule the branch migration as its own project.
What remains unverified
All Cloudflare coverage statements here, including Free-plan blocking, rule actions and the recommended mitigations, are vendor-reported from one post. No independent test of those rules was in the reviewed evidence. For depthfirst’s original disclosure and php.net’s per-branch ChangeLogs, this account relies on the cited secondary reporting for those documents. The community demonstration remains a localhost Docker lab, and the status of depthfirst’s withheld second bug is unresolved.
I would still start the patch-and-restart rollout now. Where a host lacks the demonstrated session-chain preconditions or closes FastCGI connections per request, that evidence changes its place in the rollout queue, not the destination. The fixed binary, running in every worker and checked against the effective configuration, is the only reviewed control that addresses the vulnerable components rather than one observable route into them.
Frequently Asked Questions
Which PHP releases fix all 17 vulnerabilities?
Not every release carries the full set: 8.5.13 and 8.6.0 fix all 17, while 8.4.28 fixes 14 and 8.2.36 and 8.3.37 fix 13 each where branches are unaffected, per Precursor’s tracker.
What configuration override can disable the safer default after upgrading?
The fixed releases turn session.upload_progress.enabled off by default, but an explicit session.upload_progress.enabled=1 in php.ini, an FPM pool or a virtual host overrides that safer default. A fleet can therefore run the patched binary while retaining the upload-progress entry path.
Join the discussion
Share a useful perspective or ask a question about this article.