On 17 October 2026, something tried to exploit a PHP vulnerability that did not publicly exist yet.
Amazon's threat intelligence team logged the attempt in MadPot, its honeypot network, from an anonymised address. The patch would not ship for another four days. Amazon has said it is unclear how the actor obtained exploit details during the embargo.
That is the uncomfortable opening to this story. The more interesting part comes later, and it is not about the attackers.
Part 1: the flaw underneath the application
PHP2Shell is the name given to a family of memory-safety defects in PHP itself, published on 21 October 2026 and tracked as two identifiers:
- CVE-2026-109618, described by the PHP Foundation as remote code execution via out-of-memory corruption
- CVE-2026-109619, remote code execution via stale global state
Both carry a CVSS score of 9.2.
The mechanism matters, because it explains why ordinary defences do not help. PHP has a memory limit. If a request hits that limit at precisely the wrong moment, PHP can be left holding partially initialised internal data and later treat it as valid. The second flaw is a cousin: a fatal error at the wrong point can leave internal state pointing at memory belonging to a previous request, so a later request on the same worker reuses memory that was already freed.
The PHP Foundation's advisory is specific about the consequence: an unauthenticated request can trigger this on Apache with mod_php before any application code runs, with remote code execution demonstrated under default settings. RCE was also confirmed on PHP-FPM through application-level code paths.
Read that once more. Before any application code runs.
This is the part that should change what you do this week. Your framework is not in the path. Your input validation is not in the path. The official PHP2Shell site makes the point bluntly: a simple Hello World application running on an affected PHP version can be affected.
Wiz researchers demonstrated pre-authentication remote code execution against the current container images of WordPress, Drupal and Joomla. Those were chosen as representative targets, not because they are uniquely weak. The vulnerable code is in PHP.
The second flaw also permits application-agnostic memory disclosure, which means an attacker can read what is sitting in a worker's memory without executing code at all. Passwords, tokens, API keys, certificates. Whatever that process touched.
Am I affected?
| PHP branch | Status as of Oct 2026 | First fixed release | Action |
|---|---|---|---|
| 8.5 | Actively supported | 8.5.12 | Update and restart workers |
| 8.4 | Actively supported | 8.4.27 | Update and restart workers |
| 8.3 | Security fixes only | 8.3.36 | Update and restart workers |
| 8.2 | Security fixes only until 31 Dec 2026 | 8.2.35 | Update now, plan the migration |
| 8.1 and earlier | Out of support | No patch exists | Migrate; this cannot be fixed in place |
That bottom row is the one I would worry about. Wiz reports affected releases dating back to PHP 5.4, and a great deal of shared hosting and legacy business software in India is still sitting on PHP 7.x. For those systems there is no update to apply.
Three practical notes. Updating the application does not update the runtime, so patch the server, then restart Apache, PHP-FPM and any long-running workers or the old process keeps serving. If you cannot patch immediately, the official site lists interim measures including session.upload_progress.enabled = 0, display_errors = 0 and lowering pm.max_requests, while stating clearly that these reduce exposure rather than close every variant. And Cloudflare deployed WAF rules for both CVEs roughly nine hours ahead of public disclosure, including on free plans, which helps but is not a substitute for patching.
If you want to check whether you were probed, the indicators published are unusually large or malformed PHPSESSID response cookies, unusual multipart requests followed by memory-limit errors or repeated HTTP 500s, and unexplained worker crashes. Their absence proves nothing.
Part 2: the discovery is the real story
The most remarkable part of this story is not what the researchers found. It is how long the weakness sat there, and how quickly it became exploitable once the tools changed.
These are not new lines of code. Affected releases go back to PHP 5.4, which means the underlying fragility has been shipping for over a decade in one of the most scrutinised codebases on the internet. Memory-safety bugs of this shape are genuinely hard to turn into working exploits. Knowing that a use-after-free exists is a long way from controlling it reliably across different versions, configurations and memory layouts.
That gap is what closed.
Wiz Research describes using frontier models throughout the work, and credits Gemini 4 Argon with what it calls an invaluable contribution to finding the vulnerability, exploit development, attack detection and validation of the patch. Wiz Atlas, its AI security research platform, and GPT-5.6 Sol were used throughout. The models helped read and reason about complex memory behaviour, develop and debug demonstrations, and adapt those demonstrations to different environments.
Separately and independently, OpenAI says the team Calif identified a critical flaw using GPT-5.6 Sol, as part of an initiative it calls Patch the Planet. The published credits name real people: Sagi Tzadik, Shir Tamari, Nir Ohfeld and Tomer Schwartz at Wiz, Amy Burnett at OpenAI Codex Security, and Ryan A. at Calif, working with the PHP security team and Google's Gemini team.
I want to be precise about the claim being made here, because the headline version of this story has been drifting.
No model found this alone. Skilled security researchers directed the work, chose the targets, and understood the results. What changed is the cost of the hardest part. Wiz's own framing is the striking one: frontier models enabled the team to develop and adapt working exploits across multiple applications in hours, and a single researcher completed in a week what would previously have taken a world-class team months.
That is not a claim about intelligence. It is a claim about throughput on a task that used to be a specialist bottleneck.
Part 3: what changes when this capability spreads
Be careful about which parts of this are demonstrated and which are projection. The demonstrated part is narrow and real: AI models materially accelerated exploit development and patch validation on a hard memory-safety problem, in the hands of experts.
The projection is what follows from that, and it cuts both ways.
If adapting an exploit across environments drops from months to hours, the window between a patch being published and being weaponised shrinks for everyone. Patch latency, which most businesses treat as an administrative inconvenience, becomes the main variable that decides whether you are compromised.
The defensive side is the more interesting half. The same capability points toward agents that continuously read code for defects rather than waiting for an annual review, automated regression testing when a patch lands, and the sort of deep source analysis that has historically been available only to organisations that could hire specialists.
There is an early institutional signal here too. The PHP Foundation's own post describes its handling delay as a structural failure caused by process and resourcing, and lists among its recent hires an AI Security Engineer in Residence. The people maintaining the runtime are reorganising around this.
Part 4: what you can actually use today
The honest answer has two layers, and the distance between them matters.
What is accessible now. ChatGPT and OpenAI Codex, Claude Code, and Gemini Code Assist are all available to ordinary developers and small teams. They are genuinely good at the work surrounding this incident: reading a dependency manifest, identifying an outdated runtime, explaining what a CVE means for a specific codebase, drafting the upgrade, and writing tests around the change.
A safe and useful example, which costs nothing: point an AI coding assistant at your repository and ask it to list every service, its PHP version, where that version is pinned, and which of those are on a branch that no longer receives security fixes. Then ask it what would break if each were upgraded. That is a genuine afternoon's work compressed into minutes, and it is exactly the inventory problem that makes incidents like this slow to resolve.
What is not the same thing. The research described above used frontier models inside specialist platforms such as Wiz Atlas, with sandboxed instrumentation, crash analysis tooling and experienced vulnerability researchers steering every step. A consumer subscription to a chat product is not that environment, and no amount of prompting makes it that environment. Vendors also apply safety limits to exactly this category of work.
The accurate statement is that the capability is moving from specialist research teams toward everyday engineering workflows. It has started. It has not finished.
Part 5: what I take from this
For years there was a comfortable assumption underneath a lot of security planning: some classes of vulnerability were theoretically real but practically too difficult to exploit at scale, so they sat in a backlog labelled low priority.
PHP2Shell is a direct challenge to that assumption. A weakness that had survived a decade of review in heavily audited code became a demonstrated, portable, pre-authentication exploit in weeks, because the effort required to do the hard part fell.
The same capability that makes the threat more serious is the one that gives defenders their best chance of finding these weaknesses first. Which side benefits more depends on something unglamorous: how quickly you know what you are running, and how quickly you can change it.
So the practical takeaway is not about AI at all. Go and find out what PHP version is actually running on every server you own, including the one nobody has logged into since 2023. Patch it. Restart the workers.
Then notice that the tools capable of this level of reasoning are no longer confined to frontier research labs, and ask what that means for the code you ship next year.




