The Sky Is Falling (Again): What PHP2Shell Says About AI, The Disclosure Process, and Security
Summary
Last week, PHP imploded, following the disclosure of the PHP2Shell vulnerabilities, CVE-2026-109618 and CVE-2026-109619.
Researchers leveraging AI tools discovered that some dormant vulnerabilities in PHP could be chained together into a functional exploit, which introduced PHP2Shell to the world. This exploit could allow attackers to compromise any sites that use PHP. A fix was rolled out, and chaos ensued. This story echoes what happened earlier this year, with CVE-2026-31431. Also known as CopyFail, this was another earth-shattering exploit, and one of the first major AI-assisted discoveries of a long-unnoticed but severe security issue. These two incidents have similar features, but have flawed disclosures in different ways.
During the disclosure process, technical details of the PHP2Shell vulnerabilities were posted to security forums prior to the agreed-upon disclosure date. This caused a chain of announcements and further leaks that caused panic, system shut downs, and early exploitation attempts before the actual patch was released. This brings up discussion that stresses the importance of caution during the patching of these high stress vulnerabilities.
Introduction
PHP is the backbone for many websites and is used by over 70% of the internet. It’s often used by website servers to build the webpage the user interacts with. PHP is used in other contexts as well, like load balancers or embedded devices, but websites are its most common use case. The flaws leading to the eventual PHP2Shell vulnerability have been known since 2024, but AI tooling provided the spark that led to an exploit being developed. Large website building platforms, like Wikimedia and WordPress, are at risk. The flaw was inherent to PHP, allowing even the most secure sites using affected versions to be vulnerable. The exploit could allow an attacker to execute code on affected servers, read sensitive information, or covertly steal data. These vulnerabilities, labeled as CVE-2026-109618 and CVE-2026-109619, have a staggering 9.2/10 Common Vulnerability Scoring System(CVSS) severity score. The severity was largely due to the potential for remote code execution (RCE), where attackers could execute code on the victim’s servers. RCE is the most severe an exploit can get. Not only was RCE a concern, but data exfiltration was still possible without it, giving attackers a way to steal data that was less likely to be detected. The severity of these issues and turbulent disclosure process caused widespread panic.
The PHP Foundation worked with the researchers discovering the vulnerability to create a patch. They distributed this information to other vendors so everyone could prepare and launch fixes at the same time. Once rumors of the vulnerabilities started circulating, PHP elected to announce the update on October 20th, warning users of a critical vulnerability and telling them when to expect the patch. They pushed the complete public disclosure to be a day sooner than initially planned. The pre-announcement caused a frenzy. The lack of widespread disclosure led to many administrators taking systems offline as a precaution. In the meantime, attackers were using the leaked vulnerabilities as guidelines to create exploits. Then there were the additional leaks. One was a proof of concept video of the exploit, showing use of the exploit and limited details of vulnerabilities behind it. While this video was deleted, being online at all meant that people saw it, saved it, and continued to distribute it. The other leak was an early rollout of the patch, leaving all of those that followed the embargo at a disadvantage. It is possible for would-be attackers to build exploits using comparisons between pre and post patch builds. Amazon noticed signs of potential exploit attempts as early as October 17th. Other proof of concept exploitation attempts were detected in the hours leading up to the lifting of the embargo. Once the embargo was lifted, attempts increased into the in the course of two hours. CISA added it to their list of s with guidance on how to address the issue.
The use of AI to accelerate exploit development and uncover long unexploited opportunities is indicative of what the security community can expect going forward. These hidden issues in large systems need to be handled very carefully, with releases coordinated with caution. Once a vulnerability is known or a patch is leaked, attackers begin their own investigations into it. AI tools help speed up an attacker’s ability to use vulnerabilities once information is released. A similar ordeal played out earlier this year, with a flawed disclosure process that had its own unique problems.
Looking Back At CopyFail
CopyFail was a high security exploit discovered earlier this year. Discovered in March and disclosed April 29th, it sent ripples across the security community. It had a CVSS severity score of 7.8/10, which is considered high. The exploit was built on a flaw that originated in 2017 and impacted Linux distributions since. It allows local users on the system to to “write 4 controlled bytes into the page cache of any readable file[…] to gain root”. Root on Linux systems is the highest level of privilege one can attain. This was a potential problem in many environments, as many sand box systems, cloud services, development environments, clusters and more are built on Linux. A large number of systems were at risk. Microsoft observed in-the-wild use in early May but it was largely in relation to proof-of-concept testing.
The CopyFail exploit was identified by a researcher and reported to the Linux kernel security team. This early reporting allowed for a kernel patch to be in motion before the full disclosure of the vulnerability. The problem came in with widely distributing the patch. There are many variations of Linux, and each of them is managed by different people. This creates a huge web of hundreds of individuals who need to know about the problem in order to fix it. It’s unlikely every one of them keeps it entirely secret, especially while planning ways to roll out a patch and inform their user base. Coordinating such a large group across the world is incredibly difficult, and timing is everything. If the timing is off, people are left at risk. If word gets out early, it becomes a game of who knows what and how much time there actually is before attackers catch on.
Comparing PHP2Shell and CopyFail
PHP2Shell and CopyFail feel very similar. Both are high severity vulnerabilities of widely-used systems and were discovered with AI assistance. They were built upon weaknesses that had been present for years, but AI accelerated the discovery and development of a usable exploit. Both exploits caused a frenzy when announced due to their severity, with potential for massive control over the affected system. They had messy disclosures, surrounded by panic as people realized how dangerous these vulnerabilities were. This chaos led to widespread awareness, but drew the attention of would-be attackers. There are likely plenty of similar vulnerabilities on equally critical systems, waiting to be discovered, and disclosing them properly will inform their impact.
What To Do Going Forward
These disclosure scenarios are signs of what to expect in the future. AI is speeding up vulnerability discovery and exploit development. AI excels with vulnerability discovery since it’s built around pattern matching and can be leveraged to augment security research use cases. Attackers are already leveraging AI in a variety of ways, using it to coordinate entire campaigns and develop exploits. They’re also keeping an eye on researchers, so disclosures must be carefully coordinated. There are many new tools available to attackers that they can use to make a disaster out of messy disclosures.
The leak of PHP2Shells technical details early was dangerous. It could have easily spiraled into massive damage by giving attackers advanced information with which to build their own exploits, before defenders were able to patch their own systems. If everyone operates on the same timeline, it gives maintainers time to prepare and launch the required patches, before attackers develop their own exploits with the released information. If information leaks out, attackers can take advantage of it to focus their own research. Even just the news that there is a vulnerability that is severe can be enough to lead to unwanted attention. Independent discovery of the bug by a malicious actor can defeat the purpose of the embargo, and drawing attention to an undisclosed issue increases that likelihood. This ruins the point of coordinating disclosure timing. Hopefully, going forward, there will be more diligence in the timing of information release to give defenders the best chance of patching before being attacked.
One of the key factors to remember is that it is very difficult to keep secrets amongst a large group of people. PHP2Shell was being handled across multiple groups in multiple organizations. They were collaborative and worked together for their solutions, but breakdowns in that communication may have been responsible for the leaks. Making sure everyone involved understands the reasoning behind disclosure decisions and ensuring all goals are properly communicated may help mitigate similar issues. It’s difficult to overcome human behavior, keeping secrets is not easy. There must be active effort into maintaining order to help ensure safety. Including disclosure plans in crisis planning may help prepare everyone for when high severity bugs are discovered.
Having a multi-layered approach to security helps mitigate the effects of attacks like this. Using a variety of systems and defenses, instead of relying on only one thing, can help mitigate damage. In the case of PHP2Shell, sites that heavily relied on it were more at risk, where sites that used it for some features and other systems for other features had less to worry about. Minimizing the overuse of one thing helps prevent cascading damages that destroy everything. Multiple layers of defensive tools helps in covering the weaknesses each individual tool may have.
Conclusion
AI-assisted security research has greatly accelerated the pace of vulnerability identification and disclosure. The severity of the vulnerabilities and chaos surrounding their disclosure have demonstrated that the security world needs to be paying attention. Carefully planned communication and timely rollouts of solutions are critical in preventing an actual doomsday scenario. To help mitigate the risk to you and your organization, ReversingLabs recommends scanning compiled binaries and the full install package of any software deployed in your environment.