OpenAI Says Its GPT-5.6 Sol Model Helped Find The PHP2Shell Flaws – How Does AI-Led Bug Hunting Change Open-Source Security?

The robots are finding the bugs now, but somebody still has to fix them, and that somebody is often a small team of volunteers.

A great example popped up this week with PHP2Shell, two severe security bugs in a server-side language powering huge portions of the web. While both flaws are now patched in supported versions, OpenAI said its GPT-5.6 Sol model found them, as well as forecasting that we’ll see “similar findings to increase sharply” in the near future.

In plain terms, PHP2Shell is a pair of vulnerabilities, CVE-2026-109618 and CVE-2026-109619, in the memory and error-handling code PHP uses during ordinary web requests. According to the PHP Foundation, an unauthenticated attacker could corrupt or reuse memory under the right conditions and end up running their own code on the server, which it says was demonstrated on Apache with default settings.

It’d be easy to chalk this up as an OpenAI exclusive, but plenty of others got their hands dirty. Calif.io, Wiz, OpenAI and Quarkslab all took part in the disclosure, with the PHP Foundation also crediting OSTIF, its own security team and volunteers.

OpenAI pitched the whole episode as proof that frontier AI can spot security holes before hackers do, though the developers behind PHP tell a more nuanced story.

 

Months Of Work Before The Fixes Landed

 

The actual discovery happened way before anyone made a public announcement.

According to the PHP Foundation, Calif submitted its first report back on 11 August, Wiz flagged overlapping issues on 24 August, and OpenAI independently reported the exact same root cause on 18 September. Patches only landed on 21 October, across PHP versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35.

That patch release arrived a day earlier than planned. The Foundation says CVE IDs started leaking on forums and social media around 15:00 UTC on 20 October, bundled with a wrong release time and an incorrect CVSS score of 10.0 rather than the actual 9.2, forcing an emergency schedule shift.

There were slip-ups too: Calif accidentally published and later deleted a proof-of-concept video, and Chainguard published fixed container images more than six hours before the coordinated release, according to the Foundation.

Amazon Web Services’ threat intelligence team also says its honeypot telemetry spotted an exploitation attempt dated 17 October, four days before public disclosure, which is AWS’s claim based on its own data.

 

Who Pays When The Bug Reports Pile Up?

 

The PHP Foundation says AI-assisted research is already raising report volumes. Since 1 July, its security team has published 15 CVEs and worked through more than 100 invalid reports, duplicated or couldn’t be reproduced. AI helps validate reports and group those with a shared root cause, it says, but a person still makes every judgement and works on every patch.

Fixing PHP2Shell became the security team’s top priority from 11 August, the Foundation adds, and it couldn’t have had people working full time and through weekends without Foundation funding. The fix itself was far from simple. It targeted memory and error-handling code executed on nearly every request. That left the team balancing backward compatibility, denial-of-service risks and third-party extensions.

The fixes reportedly cover only PHP 8.2 and newer, which PwC says leaves PHP 8.1 and older without an upstream fix and reliant on Linux distributions or long-term-support vendors for any backported patches.

If the Foundation’s report numbers are any guide, the workload on the receiving end is already growing. So we asked experts: how does AI-led vulnerability discovery change the workload, costs and disclosure process for open-source maintainers?

 

 

Our Experts

 

• Mudita Khurana, Staff Security Engineer at Airbnb
• Mugees Ahmed, Founder of DistroCode
• Rodion Sorokin, Co-Founder and CTO of The Capital Index
• Roman Martynenko, Founding Full-Stack Engineer at Henry AI
• Rajeshkumar Rajaseakaran Nair, Test Specialist at IBM Corporation
• T.J. Marlin, Founder and CEO of Guardrail Technologies
• Kadan Stadelmann, Founder and Former CTO of Komodo Platform
• Mandy Andress, CISO at Elastic

 

 

Mudita Khurana, Staff Security Engineer, Airbnb

 

 

“AI-led bug hunting can reduce the cost of finding complex vulnerabilities, but moves much of that cost to open-source maintainers. The PHP Foundation says it worked through more than 100 reports since July that were invalid, duplicates or could not be reproduced. Each report still needs a person to reproduce the issue, understand its impact, design a fix that doesn’t break existing software and coordinate updates with distributions and hosting providers.

“It also makes coordinated disclosure harder because maintainers may receive several AI-generated reports describing different symptoms of the same underlying flaw, while exploit development becomes faster. AI security programmes should provide a reproducible test case, a human researcher who can answer questions and enough time for downstream providers to prepare.”

 

 

Mugees Ahmed, Founder, DistroCode

 

 

“We build WordPress plugins and maintain client sites, so we see both ends of this. AI makes bugs cheaper to find, not to fix. Maintainers will get more reports and the same hours, and someone still has to reproduce each one, rule out a false alarm, then patch and test.

“Embargoes get harder to hold, because more groups find the same bug. Fortune reports that Calif and Wiz, both using AI, found parts of PHP2Shell independently in August. This time the CVE IDs leaked and PHP released a day early.”

 

 

Rodion Sorokin, Co-Founder and CTO, The Capital Index

 

 

“AI has made finding bugs cheap. Fixing them is still expensive, and the bill lands on volunteers. PHP underpins a huge share of the web, yet the PHP Foundation spent $784,376 last year and funded 11 developers. OpenAI expects similar findings to ‘increase sharply’. Maintainer capacity doesn’t.

“Disclosure is straining too. Coordinated disclosure assumes secrets keep. Here, CVE IDs leaked and PHP shipped a day early. The fair rule: if you use AI to find a bug, bring a tested patch and funding, not just a report. Otherwise maintainers get a firehose.”

 

 

Roman Martynenko, Founding Full-Stack Engineer, Henry AI

 

 

 

“AI vulnerability discovery changes the economics of maintenance before it changes the technology. Finding a flaw is getting cheaper and faster. Proving it is real, writing a fix that does not break existing users, testing it across versions, and disclosing it in the right order is still human work, and most open-source projects have very little of it to spare.

“That gap is the risk. If AI systems surface serious flaws at a much higher rate, maintainers will spend less time hunting bugs and more time triaging, validating and coordinating releases under pressure. Well-funded projects can absorb that. A small project run by volunteers cannot, even though its code may sit inside thousands of production systems.

“The answer is not to slow discovery down. It is to fund and staff the unglamorous part: triage, patch testing, and clear disclosure timelines. Discovery is scaling. Maintenance has to scale with it, or the backlog becomes the vulnerability.”

 

 

Rajeshkumar Rajaseakaran Nair, Test Specialist, IBM Corporation

 

 

“AI-led vulnerability discovery is changing the economics of open-source security. The bottleneck is no longer simply finding flaws. It’s validating them.

“AI can surface potential weaknesses at a scale manual review cannot easily match, but findings still need reproducibility checks, impact assessment, remediation and regression testing before release. For volunteer maintainers, that can mean more unpaid work, especially when AI-generated reports contain false positives or duplicates.

“Disclosure becomes harder, too. The reported early release of PHP2Shell patches after CVE identifiers circulated illustrates how quickly maintainers can face pressure when vulnerability information spreads before coordinated disclosure is complete.

“AI-generated vulnerability reports should include reproducible evidence and, where feasible, regression tests so maintainers can verify findings rather than investigate them from scratch. Organisations developing AI discovery tools should also help fund vulnerability triage, not just discovery.

“Faster discovery only improves security when maintainers can keep pace.”

 

 

T.J. Marlin, Founder and CEO, Guardrail Technologies

 

 

“Open-source maintainers are facing a workload most projects were never built to handle. AI is finding flaws in code that has been around for decades, but the teams responsible for fixing them are often small groups of volunteers. Every new report takes time to assess, and some AI-generated findings will turn out to be noise.

“The bigger concern is the shrinking window to respond. Finding a flaw is only the first step. Maintainers still need to verify it, prepare patches, backport fixes and alert users. AI is speeding up discovery, but the rest of the process has not caught up. If details leak before fixes are ready, the people relying on the software are left exposed. PHP2Shell is a reminder of how quickly disclosure can become a problem when information gets out before maintainers are ready.”

 

 

Kadan Stadelmann, Founder and Former CTO, Komodo Platform

 

 

“Maintainers are sifting through more bug reports than ever before as a result of AI-enabled sniffing. Many reports are gone through by humans, and determined to be false. Some reports fall through the cracks and do not get processed either in time or at all.

“The costs can be enormous, as in the PHP2Shell case. Resources from all over the planet, including nation-states and enterprises, were deployed to organise a fix and announcement. Patches are still ongoing. More bug reports are coming in, and those require humans to verify the issues.

“The disclosure remains overseen by humans, which leaves gaps open. There is currently a race between defenders and attackers sniffing for vulnerabilities. This is only the beginning.
“In the case of PHP2Shell, reports about the issue surfaced before the PHP Foundation, which took prompting due to the Foundation’s governance structure.

“It began with AI building upon the work of human cybersecurity researchers, and going where it was simply not practical or feasible for them to go.”

 

 

Mandy Andress, CISO, Elastic

 

 

“Over the past couple of days, several significant vulnerabilities have been disclosed, many found by AI. If it feels like the pace has changed, it has. It is not surprising; we anticipated this, but we were not fully prepared.

“As a security leader, you’re feeling this in your team. Patch queues grow, urgent advisories pile up, and the same people are asked to treat every one as a fire drill. That’s a path to vulnerability fatigue and burnout, and burned-out teams miss the things that matter. Pace yourself and your people, and be explicit that you can’t chase every CVE at the same speed.

“This is a good moment to look past the patch queue. If your exposure depends on patching faster than attackers can exploit, the architecture is working against you. Use this period to ask which fundamentals would make vulnerabilities matter less: stronger segmentation, reduced attack surface, least privilege, tighter identity controls, better detection and response for when something does get through, etc.

“Patching will always matter, but resilience is how we get through this stretch without wearing out the people defending us.”