Wiz Research reported a set of remote code execution flaws it calls PHP2Shell to the PHP security team on August 24, 2026. It was not the first report: the PHP Foundation says Calif reported a remote code execution with the same root cause on August 11, and OpenAI reported it independently. PHP shipped the fixes on October 21, 2026, 58 days after the Wiz report.
The attackers' side of the story ran in hours. Recorded Future's timeline (opens in new tab) of October 21, in UTC: an archive of leaked vulnerability information was first submitted to VirusTotal at 02:29, Patchstack reported notifications of multiple active exploitation incidents from hosting partners by 07:34, and PHP released the fixes at 13:52, eight minutes ahead of its announced 14:00 slot. For roughly six hours, exploitation was being reported and there was no PHP release to install. Recorded Future adds that the available reporting "does not establish the full extent of successful exploitation or confirmed compromises."
AI made the finding and the proof cheap. Two parts stayed slow: deciding what the fix should be, and installing it on the branches real servers run. A patch deadline keyed to a severity score assumes both happen on time. Neither held here. The controls worth taking from this are aimed at those two slow parts, and each one ends in a date or a name.
AI found it. The branch decides whether you can fix it.
Wiz says it found the first bug on August 19 using Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol. Its write-up (opens in new tab) states that "Every PHP version released in the last decade is affected by at least one of these vulnerabilities," going back to PHP 5.4. In plain terms, running out of memory partway through a request can leave internal data half built, and the cleanup code then trusts it. State from one request can also outlive its memory and be picked up by the next request that worker serves. PHP groups them under CVE-2026-109618 and CVE-2026-109619, both rated 9.2 on CVSS v4.0.
The line in PHP's advisory (opens in new tab) that should change how a fintech scopes this is the risk statement: "all PHP installations serving web requests." PHP rates the risk highest for Apache with mod_php, where a demonstrated exploit lets an unauthenticated request trigger the attack before any application code runs. Careful application code does not help when the flaw sits underneath it. AI pulling hundreds of bugs out of Firefox was a warning about your own repositories. PHP2Shell puts that pressure on the runtime under them, where most firms control the fix least.
Then the branches. PHP shipped 8.5.12, 8.4.27, 8.3.36 and 8.2.35, and the advisory is direct about the rest: "PHP 8.1 and older no longer receive updates from the PHP project." Given the severity, the security team published minimal source backports for 7.4, 8.0 and 8.1. That is real help, and it is still not a release. Someone has to build and ship a patched interpreter, buy one from a support vendor, or migrate to a supported branch with its own test cycle.
The scale of that problem comes from scanners. Recorded Future relays counts from Modat: 4,384,302 PHP services on 1,886,612 unique hosts, and "About 79% of services with a readable version run a branch PHP no longer supports." PHP 7.4, end of life since 2022, accounts for 878,761 services, more than any supported branch. Hold the percentage loosely. It counts services that announce a version, and Recorded Future, which pairs it with a VirusTotal scan, notes that neither scan counts confirmed vulnerable hosts. Both point in one direction, and that is enough to plan on.
In a regulated fintech, I would expect the exposed PHP to sit outside the core platform. These are the places I would look first:
- The agency-hosted marketing site, often WordPress and often missing from the asset inventory because it was bought as a service
- A partner or broker portal built years ago, still taking logins, owned by whichever team inherited it
- The admin console on a vendor appliance, running whatever PHP the vendor shipped and patched on the vendor's schedule
- A container base image rebuilt by a pipeline with no named owner, carrying whatever interpreter the upstream image shipped
Recorded Future's guidance says to include PHP bundled in third-party apps, containers and managed services. So the inventory has to reach those too, and name the PHP branch and the owner of every internet-facing host. Next to it sits a retirement plan for each end-of-life branch, with a target date and a funded migration. An open-ended promise to upgrade later is not a plan I would accept for an internet-facing system its project no longer fixes.
Patched was not the end state
PHP's advisory lists interim measures for teams that cannot update, then states what they are worth: they "reduce exposure to specific attack paths but do not make an installation safe."
Edge filtering was no different. Cloudflare wrote (opens in new tab) that the vulnerability "was leaked before the official release date" and that it observed attack traffic targeting its customers. Its rules targeted known exploit attempts, and its October 23 update on the related CVE-2026-115560 is plain about where a rule stops: sites that keep PHP's default name for the upload progress field are covered, and "Customers who have changed session.upload_progress.name are not covered by the WAF rule."
A WAF rule written for one disclosed flaw is a virtual patch. It blocks the request patterns its author anticipated, and its coverage stops at the edge of those patterns. Cloudflare's own advice is to patch anyway, so a risk acceptance that cites the WAF has to state whether the site matches the pattern.
Then a fully patched server turned out to have one more narrow path. On October 23, PHP published a notice about CVE-2026-115560, which Cloudflare says "affects fully patched PHP releases, including 8.5.12." PHP calls it an unlikely vector that needs an application to switch its session serializer at runtime, and suggests keeping session.upload_progress.enabled=0 until the next regular bugfix release, a setting it had already listed as a mitigation on October 21. So for any application that might switch serializers at runtime, the remediation record is a version and a configuration line. A version check alone cannot show the second half.
The 8.2.35 note carries one more date: "PHP 8.2 reaches end of life on 31 December 2026 and will receive no further security fixes after that date." A team that moved to 8.2.35 on October 21 bought about ten weeks of project support. Without a retirement plan for 8.2 already in motion, that clean patch report is a countdown.
Deciding the fix was the slow step
The most useful document in this episode is the PHP Foundation's account (opens in new tab) of why a root cause its team first saw in 2024 stayed unfixed until October 2026, because regulated firms reproduce that failure inside their own exception processes. That root cause is the out-of-memory (OOM) problem behind CVE-2026-109618. At that point, the Foundation writes, "a potential exploit for the OOM problem was theorized, but no concrete vectors were found." Every repair on the table carried a cost: "backward compatibility breaks, new denial-of-service potential, or requiring changes to many third-party extensions." The team, "operating on consensus, struggled to choose between these" and "feared the repercussions of shipping the wrong choice." Work on a fix stalled and restarted over months. The Foundation's verdict on that stall is unusually candid: "the delay was a structural failure, caused by both the security team's decision process and its lack of resources."
The Foundation says the reports from Calif, OpenAI and Wiz made the full impact clear, and from August 11 the fix became its top priority. The remedy it describes for the stall is a governance change: the security team has started moving so that decisions can be made by majority, with compromise encouraged and leadership able to steer a stalled case. The post also credits AI-assisted research with bringing the team far more reports than before.
Swap in your own nouns. The hard fix breaks an integration, risks an outage, or needs a third party to change something. Nobody wants to own the wrong call, so the decision waits for agreement and the exposure stays open while it waits. More findings make that wait more expensive. So does an attacker who needs only hours once details leak.
I read a risk acceptance the way the CISSP, CISA, CISM and CRISC material trains you to read any control: who owns it, and when it gets tested again. For a runtime flaw like this one, here is what a usable acceptance carries.
- An expiry date set by exposure. An internet-facing PHP 8.1 host gets weeks. An internal host behind authentication can get longer, and the record says why.
- A named decider. One person, by name and role, with authority to accept the risk for the business that owns the asset. A committee can advise. It cannot be the decider.
- The compensating controls and their gaps. If the WAF rule is listed, the record states which exploit paths the rule covers and whether the site matches its assumptions, such as the default upload progress name for CVE-2026-115560.
- The conditions that void it early. A listing in CISA's Known Exploited Vulnerabilities (KEV) catalog or a change in exposure ends the acceptance and restarts the clock.
Then report exception aging the way you report open findings, by age band and by exposure, with the decider's name on every line. When an acceptance expires, it escalates on that date to the next level of authority, whether or not the group has reached agreement. That is the Foundation's majority rule, translated into a risk register.
Put the clock on exposure
A typical patch policy sets the deadline by severity score, counted from the day a fix is released. That clock failed twice here. Exploitation was reported before there was a fix to count from, and for every branch at 8.1 and older the project release never arrives. A 9.2 on an unsupported branch then sits in the queue with a deadline that cannot be met.
CISA's Binding Operational Directive 26-04 (opens in new tab) (BOD 26-04), issued June 10, 2026, sets urgency from four variables instead: whether the vulnerable asset is publicly exposed, whether the CVE is in the KEV catalog, whether an adversary can automate every step of the exploit, and whether exploitation gives partial or total control of the asset. Severity, in the CVSS sense, survives only as the last of the four. The timeline begins when CISA adds the CVE to the catalog or when the agency finds it on an asset, "Whichever event occurs first." It revokes BOD 22-01, and it names the pressure behind the change: attackers' "use of AI may further narrow the time defenders have to react between patch release and possible exploitation."
BOD 26-04 binds federal civilian executive branch agencies. It does not bind a fintech, so the honest use is to borrow its logic. Run these two CVEs through the four variables and the internet-facing partner portal on 7.4 goes ahead of an internal 8.3 host with the same score. The directive also names a move that does not wait for a release: taking a system off the internet changes its exposure value and pushes the required timeline back. For the vendor appliance still waiting on its vendor, that decision belongs to the firm.
The four variables also give time back. The directive is explicit about "deferring action against low-risk vulnerabilities," which is how a team pays for moving faster on the exposed ones. Exploit automation is the variable AI moves first, since what AI changes for attackers is mostly the cost and scale of steps that were already routine.
CISA added both CVEs to the KEV catalog on October 26, 2026, citing BOD 26-04 (opens in new tab). On October 27, Wiz wrote (opens in new tab) that it had "not yet observed successful end-to-end PHP2Shell exploitation" in its own telemetry, while it had seen "a limited number of exploitation attempts." That post calls researcher Adel Karimi's honeypot data "the most robust public evidence" of code execution it had seen, and names his results as a likely contributing factor in CISA's decision to list PHP2Shell. For the clock, the listing is what shortens it, and it came five days after disclosure. Wiz also counted less than 53% of initially affected organizations still vulnerable after one week.
The question I would bring to an exception register is the one the PHP Foundation had to answer about its own process, asked entry by entry: who settles a decision that has stalled. An exception with no expiry date and no named decider is a decision nobody made.