Capabilities

AI Found Two Critical PHP Flaws: What It Shows About Machine Vulnerability Discovery

Models from OpenAI and Google helped researchers find bugs that sat in PHP for years. Here is what happened, what is disputed and what it says about AI security work.

Claire Nakamura

Technical Writer, Frontier AI Coverage

Published 8 min read
High-resolution close-up of HTML code displayed on a computer screen, perfect for technology themes.
In this story 10 sections

AI models helped researchers find two critical flaws in PHP, the language behind roughly 70% of websites. The PHP team shipped emergency fixes on October 21, 2026. Both bugs allow remote code execution without a login. The case shows AI can find deep bugs fast, but it also shows how quickly attackers move once a patch appears.

This explainer is for developers, site owners and anyone who wants to know what machine vulnerability discovery really looks like. It covers what happened, which facts are firm and which are disputed, and what to do about it. Emergent Wire covers AI capabilities, and this story is one of the clearest real-world tests yet of AI doing security research at scale.

What Happened With the PHP Flaws?

Close-up of a smartphone wrapped in a chain with a padlock, symbolizing strong security.

PHP's developers fixed two critical vulnerabilities that could let attackers take over servers without logging in. ABC News reports both are rated 9.2 out of 10 for danger. The PHP team released the updates on Wednesday, October 21, 2026, and urged all users to update immediately, according to ABC News.

The reach is large. W3Techs estimates PHP runs roughly 70% to 75% of websites whose language can be identified. PHP also underpins WordPress, Drupal, Joomla and Magento sites, so a flaw in the language reaches a huge slice of the web.

The disclosure was moved up. ABC News says the planned announcement came sooner after the vulnerabilities began circulating on forums and social media. That detail matters because it shortened the time defenders had to prepare.

How Did AI Find the Bugs?

Several groups used AI models in different ways. Reuters reports that the security firm Calif found the flaws with a large language model on August 1. The firm Wiz found them using its autonomous AI tool, Wiz Atlas, together with Google's Gemini 4 Argon and OpenAI's GPT-5.6 Sol, according to the Reuters report.

Wiz said that without AI the manual work would be difficult to implement and even harder to scale. It also said the flaws had stayed hidden for more than a decade.

OpenAI says its models helped discover the bugs and that it led a coalition of maintainers, researchers and security providers to prepare fixes before disclosure. The PHP Foundation had entered PHP into OpenAI's Patch the Planet program, and a security audit began in July, ABC News reports. Work on the fix began August 11.

Why Did Bugs Like These Stay Hidden for Years?

Large, old codebases hide problems in places no human reviewer has time to check. PHP has been developed for decades, and its core handles parsing, memory and file uploads in complex ways. A bug can sit in a rarely used path for years without any test hitting it.

AI changes the economics of that search. A model can read huge amounts of code, form a guess about how a function could be abused and test it faster than a person. It does not get bored, and it can try thousands of paths in parallel.

ABC News adds that a potential exploit was theorized as early as 2024, but no concrete vectors were found, and work on a fix stalled and restarted over months. The PHP Foundation called the delay a structural failure. In other words, the weakness was hinted at long before it was proven.

What Do Sources Agree On, and What Is Disputed?

A cybersecurity expert monitors multiple screens, focused on data protection in a dark room.

Sources agree on the core facts. The flaws are critical, AI models were involved, and patches shipped on October 21. They differ on details such as how widely the bugs are being exploited, who found what first, and how many servers remain exposed.

The table below sets out the main points of agreement and disagreement.

What Do Sources Agree On, and What Is Disputed?
QuestionWhat is reportedStatus
SeverityTwo critical flaws, 9.2 of 10Consistent
Patch dateOctober 21, 2026Consistent
Active attacksAmazon says broad exploitation; CISA had not observed any at publicationDisputed
Unpatched hostsAbout 1.5 million of 2.6 million scanned, per CensysSingle source
Discovery creditCalif, Wiz and OpenAI each describe a roleOverlapping

On exploitation, Amazon Threat Intelligence says it saw a hacking attempt against a decoy system on October 17, four days before the fix. It calls the exploit PHP2Shell and says it is now being broadly exploited. CISA said it had not observed exploitation when its statement was published, and ABC News does not reconcile the two.

How Big Was the Fallout?

Large enough to take real services offline. Reuters reports that the Netherlands' National Cyber Security Centre went offline as a precaution, along with the Port of Rotterdam, Overheid.nl, several hospital sites and the websites of the five largest political parties. Denmark's Kombit shut down several services, Finland's Valtori took government public network services offline, and France's Interior Ministry placed all its websites in preventive maintenance.

The effects reached beyond governments. Wikipedia's operator took down unspecified systems, and the Internet Archive briefly went offline. Censys data cited by Reuters shows about 1.5 million of 2.6 million hosts running a PHP version with no fix, roughly 58%.

Some platforms were safe. Automattic says WordPress.com and Tumblr users are not at risk. Small PHP sites on shared hosting face higher risk, according to the same report. Emergent Wire's editors note that these figures come from a single snapshot and will change as sites patch.

What Does This Say About AI and Cybersecurity?

A classic MS-DOS terminal screen displayed on a laptop keyboard with vivid illumination.

It shows AI can produce real, high-severity findings, not just noise. OpenAI called the bug one of the most impactful vulnerabilities in recent memory. That is a strong claim, and it comes from a party with a stake in AI's reputation, so readers should weigh it accordingly.

It also shows a dual-use problem. The same models that help defenders find a bug can help attackers find the next one. Our earlier report on ARTEX and the ConnectWise breach covered an open-source AI attack framework, and the PHP case is the defensive side of the same capability.

The race is now about speed. If AI shortens the time from bug to exploit, defenders need to shorten the time from patch to deployment. Amazon advises applying critical fixes to internet-facing systems within hours, not days or weeks.

Can AI Coding Tools Prevent Bugs Like These?

Partly. AI tools can review code and flag risky patterns before release, which is one way to catch bugs earlier. Our guide to where AI code review actually works covers where those tools help and where they miss problems.

They are not a cure-all. A bug in a mature, complex system often needs deep context to find, and AI tools can also write insecure code. Our look at AI coding agents in 2026 shows how far unattended agents can go and where human review still matters.

The practical answer is layers. Use AI review, human review, fuzzing and a fast patch process together. No single tool has to be perfect if the system catches what each one misses.

How Can Teams Build a Faster Patch Process?

Speed comes from preparation, not heroics. Teams that patch within hours usually know exactly where each runtime lives, have a tested way to roll out an update and can roll it back if something breaks. Teams that do not have that map spend the first day just finding servers.

Start with an inventory. List every application, container image and managed service that includes PHP, and note who owns each one. A simple spreadsheet beats a perfect tool that nobody finishes.

Next, rehearse. Pick a low-risk system and run a full patch cycle on it, from test to production. The first run exposes the slow steps, such as manual approvals or missing staging environments, so you can fix them before an emergency forces the issue.

Finally, agree in advance on a rule for emergencies. For example, a critical flaw on an internet-facing system gets patched within 24 hours, with a named person who can approve it. A rule written in calm weather is easier to follow when a headline breaks.

What Should Site Owners Do Right Now?

Close-up of a woman using a smartphone and laptop with red keyboard in an indoor setting.

Update PHP. The PHP team says updating is the only complete fix. If you run a managed platform, ask your host whether it has patched, and get the answer in writing.

Follow these steps in order, starting with the fastest.

  1. Find every server and container image that runs PHP.
  2. Apply the updated PHP release to internet-facing systems first.
  3. If you cannot update at once, take the system offline until you can.
  4. Check logs for suspicious activity going back to October 17.
  5. Rotate credentials on any server that may have been exposed.

The FBI and the UK's National Cyber Security Centre recommend taking an affected system offline if an immediate update is not possible, according to ABC News. Temporary changes to memory and upload settings only reduce exposure and do not make a server safe.

A fair counterpoint is that the story is still developing. Details on exploitation, affected versions and credit may change as researchers publish more. Check the PHP project's own advisories for the current version numbers.

Conclusion

AI helped surface two critical PHP flaws that had stayed hidden for years, and the patch race that followed shows how fast the field now moves. The facts on severity and the patch date are solid, while attack volume remains disputed. If you run PHP anywhere, update today, review your logs back to October 17 and set up a process for patching internet-facing systems within hours.

What are the critical PHP vulnerabilities patched in October 2026?
PHP developers fixed two critical flaws rated 9.2 out of 10 that could allow unauthenticated remote code execution, letting attackers take over servers. The PHP team released fixes on October 21, 2026, and urged all users to update immediately. Updating PHP is the only complete fix, according to the PHP team.
How did AI help find the PHP flaws?
Reuters reports that Calif found the flaws with a large language model on August 1, while Wiz used its Wiz Atlas tool with Gemini 4 Argon and GPT-5.6 Sol. OpenAI says its models helped researchers and that it led a coalition preparing fixes before disclosure, which shows AI doing real security research.
Are the PHP flaws being exploited in the wild?
Sources disagree. Amazon Threat Intelligence says it saw a hacking attempt on October 17 and that the exploit, called PHP2Shell, is now broadly exploited. CISA said it had not observed exploitation when its statement was published. Treat the risk as active and patch internet-facing systems immediately.
What should I do if I cannot update PHP right away?
The FBI and the UK National Cyber Security Centre recommend taking the affected system offline until it can be fully mitigated. Temporary changes to memory and upload settings only reduce exposure. Also check logs for suspicious activity back to October 17 and rotate credentials on any server that may have been exposed.

Written by

Claire Nakamura

Technical Writer, Frontier AI Coverage

8+ years experience

Claire Nakamura started as a developer relations engineer at two developer-tooling companies before moving into full-time technical journalism. She spends most of her working week with API documentation, model cards, and changelogs that most outlets skip past, and she is the person on staff most likely to have actually run the eval a press release is citing. At Emergent Wire she covers developer-facing AI tooling, agent frameworks, and the practical gap between a model's stated capabilities and what shows up in a production integration.

Credentials

  • Former developer relations engineer, developer-tooling sector (6 years)
  • Hands-on tester of API-based AI tools and agent frameworks
  • Regular contributor on model card and changelog analysis