Lyron
Laptop showing program code on screen
Guide

AI is finding security flaws faster: what PHP2Shell means for SMEs

· 10 min read
By Editorial quality standard

On October 21, 2026, the PHP project fixed two critical security flaws that became known as PHP2Shell. PHP runs a large share of all websites – from a tradesperson’s WordPress site to online shops. Artificial intelligence was involved in finding the flaws, and five days after the updates the US cybersecurity agency CISA already listed both as actively exploited.

PHP2Shell is part of a development that showed up several times in the past few days: AI models are finding security flaws faster than ever – for defenders and attackers alike. Google is holding back its new model Gemini 4 Carbon for security teams because of its capabilities, Mistral has released an open model for defenders, and a study shows how openly available AI turns published fixes into attacks. For small and mid-sized businesses, one conclusion matters most: updates need to land faster.

In brief

  • PHP2Shell: Two critical flaws in PHP (CVE-2026-109618 and CVE-2026-109619) let attackers run code on web servers without logging in. The fix is PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35.
  • Actively exploited: CISA has listed both flaws as exploited since October 26. Amazon saw a first exploitation attempt four days before disclosure.
  • AI as a vulnerability finder: According to OpenAI, the security firm Calif found one of the flaws with the OpenAI model GPT-5.6 Sol; Wiz discovered both independently.
  • The patch gap is shrinking: Epoch AI shows that an openly available model can build working attacks from published fixes – for about 1,000 US dollars per Firefox version.
  • What SMEs should do: check the PHP version or ask the host, list all systems, agree on response times for critical updates and track security advisories automatically.

What is PHP2Shell?

PHP is the programming language behind a large part of the web: WordPress, many online shops, customer portals, booking and ticketing systems. On October 21, 2026, the PHP project released versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35. They close two critical flaws that let attackers run their own code on the server without logging in. Both carry a CVSS score of 9.2 out of 10; according to Germany’s Federal Office for Information Security (BSI), neither a login nor any user action is required. The name PHP2Shell deliberately echoes Log4Shell, the flaw that put systems at risk worldwide at the end of 2021.

According to PHP, every installation that serves web requests is affected – whether it runs through PHP-FPM, Apache with mod_php, php-cgi, FrankenPHP or LiteSpeed; the risk is highest with mod_php. CrowdStrike notes that it does not matter which CMS, framework or plugins run on the server. Updating WordPress or its plugins therefore does not close the flaws – only an updated PHP does.

The key dates at a glance:

  • October 17: Amazon records a first exploitation attempt in its honeypots – four days before disclosure.
  • October 21: PHP releases the updates and urges all users to update immediately. According to The Hacker News, the release came a day earlier than planned because details had leaked. The BSI issues a cybersecurity warning.
  • October 22: The BSI adds that makers of network components are also publishing advisories, because PHP is sometimes built into their products.
  • October 26: CISA adds both flaws to its catalog of known exploited vulnerabilities. US federal agencies had to patch by October 29.

PHP 8.1 and older no longer receive updates from the PHP project. Because of the severity, the security team made an exception and provided minimal fixes for PHP 7.4, 8.0 and 8.1, intended for anyone who builds PHP from source; some Linux distributions ship their own fixes. Anyone still running such an old version should treat PHP2Shell as the moment to upgrade.

How AI was involved in the discovery

According to OpenAI, the security firm Calif found one of the two flaws with the OpenAI model GPT-5.6 Sol as part of OpenAI’s “Patch the Planet” initiative. The security company Wiz discovered both flaws independently; the BSI also names Wiz and Calif as involved in the discovery together with OpenAI. OpenAI writes that, to its knowledge, AI found the flaws before attackers did, giving the PHP maintainers time to prepare fixes.

For context: how much of the find was down to AI cannot be verified from the outside – the claims come from the vendor that sells the models. In addition, BleepingComputer reports that details and sample code circulated before the release; according to the PHP security team, that code did not work. The direction, however, is clear: OpenAI expects findings like these to increase sharply in the coming months. More flaws found means more updates – and less time to install them.

The patch gap: why attackers are getting faster

Between the moment a fix becomes public and the moment it reaches users lies the so-called patch gap. A study published by Epoch AI on October 29 shows how dangerous this window has become: the openly available model GLM 5.3 from Z.ai identified which changes in Firefox’s published source code fixed security flaws and turned them into attacks on six recent Firefox versions. A crafted web page could read files from a visitor’s computer. The model also used a previously unknown flaw, which Mozilla fixed on October 27 with Firefox 159.

Once the test setup was built, each successful attack cost about 1,000 US dollars according to Epoch; the attacks were built within two days. The researchers’ conclusion: once a fix is in published code, assume attackers will use it against the current release. Firefox and Chrome now ship updates every week.

For SMEs this means the time between “update available” and “attack under way” is shrinking from weeks to days. If updates only go in at the monthly maintenance slot, critical flaws stay open for too long.

Gemini 4 Carbon and Mistral Large 4: AI models for defenders

Two announcements this week show how seriously AI vendors take the issue. On October 28, Google introduced Gemini 4 Carbon, by its own account its strongest model, with top scores for finding, exploiting and fixing vulnerabilities. Access goes first to a subset of cyber defenders, then gradually to more partners through Google’s Fairwind Program. Developers, businesses and consumers are to follow “as soon as possible”; Google gives no date.

A day later, the French vendor Mistral released the weights of Mistral Large 4. The one-trillion-parameter model is geared towards coding and defensive security work and, according to Mistral, was trained on European infrastructure. Companies can therefore run it on their own hardware – although at this size that is a job for large organizations and security providers rather than a small business.

AI models with security capabilities compared, as of October 30, 2026
ModelVendor and dateAccessSecurity results according to the source
Gemini 4 CarbonGoogle, October 28, 2026initially selected security teams only (Fairwind Program)88.2% on SEC-bench Pro (finding flaws), 54.8% on ExploitGym (exploiting), 71.0% on CWE-bench (fixing)
Mistral Large 4Mistral, October 29, 2026open weights51% on the Artificial Analysis Cyber Index
GLM 5.3Z.ai, August 2026open weightsbuilt attacks on six Firefox versions in the Epoch study

The figures are vendor claims or results of the respective study; independent tests were not available at the time of writing. None of these models is a tool SMEs need to run themselves. What matters is the direction: flaws are found faster – and exploited faster. A business that installs updates quickly therefore matters more than the strongest model.

What SMEs should do now: six steps

  1. Check the PHP version: Ask your host or service provider in writing whether every server running PHP has the fixes from October 21, 2026. Linux distributions often backport fixes to older version numbers, so the version number alone says little. On shared hosting the host usually installs the update; on your own server or VPS it is your responsibility.
  2. List every PHP system: not just the website, but also the shop, customer portal, booking and ticketing systems, intranet – and network devices that may contain PHP.
  3. Have the logs checked: Because the first exploitation attempts appeared on October 17, logs should be reviewed from that date onwards. According to CrowdStrike, an attack leaves hardly any traces: the application logs nothing unusual, and the only sign is a single memory-limit error in the web server’s error log.
  4. Agree on response times: Decide with your provider how quickly critical updates are installed – for example within 72 hours if a flaw is already being exploited. For comparison: CISA gave US federal agencies three days for PHP2Shell.
  5. Retire old versions: PHP 8.1 and older no longer receive regular updates from the PHP project. Plan the upgrade now, not when the next flaw arrives.
  6. Clarify responsibilities: Who receives security warnings, who decides, who installs, who checks afterwards? This also applies to self-hosted tools: your own n8n instance does not run on PHP, but it needs regular updates just the same – more in our comparison n8n self-hosting vs. cloud.

If you cannot update right away, PHP’s security advisory lists interim measures, such as setting session.upload_progress.enabled=0. PHP stresses, however, that only the update fully protects a server.

Track security advisories automatically: a workflow with n8n

PHP2Shell will not be the last flaw of its kind. Instead of watching news sites, the monitoring can be automated – for example with n8n, Make or Zapier. A lean workflow has four building blocks:

  • Fetch the sources: the news of the software you use (PHP offers an Atom feed, for example), the BSI’s cybersecurity warnings and CISA’s catalog of known exploited vulnerabilities, which the agency also publishes as a JSON file.
  • Match against your inventory: a simple list of your systems with software and version is enough; the workflow only reports matches that affect you.
  • Create a task with a deadline: each match creates a ticket for the person responsible; the deadline depends on the severity.
  • Remind and escalate: if the task stays open, the workflow sends reminders and informs management once the deadline has passed.

Fictional example: An online retailer with 15 employees runs its shop, website and customer portal with two hosting providers. Every morning an n8n workflow checks the advisories from PHP, the BSI and CISA, matches them against the system list and, if there is a match, creates a ticket with a 72-hour deadline. If nobody picks it up within 24 hours, a reminder goes to management. A person still installs the update – the workflow only makes sure no advisory gets lost. The scenario is made up – not a client project, no measured results.

An AI agent can also summarize the advisories and pre-sort them by urgency; the decision to update stays with a person. How to keep such workflows secure themselves is covered in our article on security for automations. Manufacturers that offer products with digital elements in the EU have had to report actively exploited vulnerabilities under the Cyber Resilience Act since September 11, 2026; our guide to the 24-hour reporting workflow shows how to organize it.

This article is practical guidance and does not replace individual security advice.

Frequently asked questions

What is PHP2Shell?

PHP2Shell is the nickname for two critical security flaws in PHP (CVE-2026-109618 and CVE-2026-109619) that were fixed on October 21, 2026. They let attackers run code on web servers without logging in. Only an update to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35, or a vendor package that contains the fixes, fully protects a server.

Is my WordPress website affected by PHP2Shell?

Possibly: the flaws are in PHP itself, not in WordPress. Updating WordPress, plugins or themes therefore does not close them. On shared hosting the host usually installs the PHP update; ask whether the fixes from October 21, 2026 are in place.

What role does AI play in PHP2Shell?

According to OpenAI, the security firm Calif found one of the two flaws with the OpenAI model GPT-5.6 Sol; Wiz discovered both independently. At the same time, a study by Epoch AI shows that openly available AI models can build attacks from published fixes. That shrinks the time between an update and an attack.

How quickly should SMEs install critical updates?

For flaws that are already being exploited, within a few days rather than weeks. The US agency CISA gave federal agencies three days for PHP2Shell. Agree on response times like these in writing with your host or service provider.

Sources

As of October 30, 2026. Facts and figures come from the sources listed; context, example and recommendations are Lyron’s.

Want to stop leaving security advisories to chance?

In a free consultation we look at which of your systems run on PHP, who updates them and how a workflow can route security advisories to the right person automatically.

Free Consultation

Share article: